Skip to content
ProductArena

AI Assistants Arena

ChatGPT vs Claude

ChatGPT logo

ChatGPT

OpenAI

Arena38/100AGENT-READY49/100BUILT-IN AI86/100
MCP✓CLI✓API~
free-tier9/100 integrity
Claude logo

Claude

Anthropic

Arena29/100AGENT-READY28/100BUILT-IN AI86/100
MCP✓CLI!API~
free-tier0/100 integrity

Claude wins · 16–14 (19 drawn)

Agenticness — how well agents can access and operate the productAgenticness

How well agents can access and operate the product

Agent access

  1. ai-native userPoint an agent at llms.txt or agent-oriented docs

    weight 2 · round to Claude
    ChatGPTpartialprobed5/10

    ChatGPT's Browser and MCP features let a user direct the agent to fetch and use third-party documentation or websites (e.g. Context7 for developer docs, browsing arbitrary URLs), which would technically allow pointing it at an llms.txt or agent-oriented doc page. The evidence pack also shows ChatGPT's own docs site publishes an llms.txt (probe-1), indicating familiarity with the convention, but there is no explicit documented feature or example of a user instructing ChatGPT to consume an llms.txt file specifically. Missing for 10: explicit product support/example for llms.txt ingestion, confirmation that browsing normalizes/parses such files as agent-context rather than generic web content, independent hands-on verification.

    • [probe] “PROBE llms.txt: HTTP 200 at https://learn.chatgpt.com/llms.txt # Codex > Build with the Codex CLI, IDE extension, and cloud automation to s…”
    • [claimed-docs] “Browser lets ChatGPT open websites, gather current information, and take action while you stay in control.”
    • [claimed-docs] “Use it to give ChatGPT or Codex access to third-party documentation, or to let it interact with developer tools like your browser or Figma.”
    • [claimed-docs] “Model Context Protocol (MCP) connects models to tools and context. Use it to give ChatGPT or Codex access to third-party documentation”
    • [claimed-docs] “to add Context7 (a free MCP server for developer documentation)”
    proof ↗⚑ flag
    Claudefullprobed7/10

    A probe confirms Claude's support site serves a working llms.txt (HTTP 200) listing topic links, and numerous individual documentation pages are available in clean .md format (e.g. claude-docs-1 through 63 all resolve as .md URLs), making the docs directly consumable by an agent. Missing for 10: confirmation that llms.txt/agent-readable docs exist on the main claude.com domain (only support.claude.com was probed), and independent/community evidence that agents actually consume these successfully rather than just first-party doc structure.

    • [probe] “PROBE llms.txt: HTTP 200 at https://support.claude.com/llms.txt # Claude Help Center > Search for answers or browse by topic ## English #…”
    • [claimed-docs] “You can prompt Claude to search through your previous conversations to find and reference relevant information in new chats.”
    • [claimed-docs] “This article explains how chat search and memory work, what Claude does and doesn’t remember, how to review and edit what’s saved, and how t…”
    • [probe] “PROBE docs-md: HTTP 404 at https://support.claude.com/en/.md”
    proof ↗⚑ flag

  2. ai-native userRun the product headlessly / in CI for automation

    weight 2 · round to ChatGPT
    ChatGPTfullclaimed7/10

    ChatGPT's Codex offers a CLI and SDK explicitly documented for headless/CI use ('Use the Codex SDK to automate coding tasks, including jobs in CI'), plus CLI flags (--search, /fast) and a bulk-scan command (npx @openai/codex-security bulk-scan) that support non-interactive automation workflows. missing for 10: independent/hands-on confirmation of actual CI pipeline runs, and more detail on authentication/headless setup specifics for CI environments.

    • [claimed-docs] “Use the Codex SDK to automate coding tasks, including jobs in CI.”
    • [claimed-docs] “Inspect code, make changes, run commands, and automate repeatable work without leaving your terminal.”
    • [claimed-docs] “Use its command-line interface (CLI) to scan repositories you own or have permission to assess, review findings over time”
    • [claimed-docs] “Use `npx @openai/codex-security bulk-scan` to review repositories in one campaign.”
    • [claimed-docs] “Use `/fast on`, `/fast off`, or `/fast status` in the CLI to change or inspect the current setting.”
    • [claimed-docs] “In the CLI, pass `--search` to fetch live results for one run”
    proof ↗⚑ flag
    Claudenone0/10

    The evidence describes Claude Code as a terminal tool and mentions CLI/print output structure, but there is no documentation or claim of headless/non-interactive execution, CI pipeline integration, or scriptable automation mode for Claude products.

    • [claimed-docs] “Build with Claude Code in your terminal, then deploy to a URL Claude can reach”
    • [claimed-docs] “The screen reader mode brings Claude Code back to the basic terminal experience: plain, sequential text with added labels and cues”
    • [community] “src/cli/print.ts is the single worst function in the codebase by every metric: 3,167 lines long, 12 levels of nesting at its deepest, ~486 b…”
    proof ↗⚑ flag

  3. ai-native userPlug MCP servers into this product so it can use their tools

    weight 3 · round to Claude
    ChatGPTfullclaimed8/10

    ChatGPT has first-party documentation for MCP support, explicitly describing configuring MCP servers (e.g., Context7) so ChatGPT/Codex can access third-party tools like documentation, browser, or Figma, with config stored in config.toml and scoping options for trusted projects, plus cross-client portability once configured. missing for 10: independent/hands-on community corroboration of MCP server usage specifically (community evidence covers plugins/browsing but not MCP directly).

    • [claimed-docs] “let it interact with developer tools like your browser or Figma”
    • [claimed-docs] “Once you configure your MCP servers, you can switch among those clients without redoing setup.”
    • [claimed-docs] “Use it to give ChatGPT or Codex access to third-party documentation, or to let it interact with developer tools like your browser or Figma.”
    • [claimed-docs] “to add Context7 (a free MCP server for developer documentation)”
    • [claimed-docs] “you can also scope MCP servers to a project with `.codex/config.toml` (trusted projects only)”
    • [claimed-docs] “Codex stores MCP configuration in `config.toml`”
    • [claimed-docs] “Model Context Protocol (MCP) connects models to tools and context. Use it to give ChatGPT or Codex access to third-party documentation”
    proof ↗⚑ flag
    Claudefullprobed9/10

    Claude documents connecting to both remote MCP servers via custom connectors and installing local MCP servers on Claude Desktop as easily as browser extensions, plus a unified directory for finding/installing connectors, letting Claude use their tools. Missing for 10: independent/hands-on verification of MCP tool usage reliability beyond vendor docs.

    • [claimed-docs] “Build your own remote MCP servers to connect with any tool.”
    • [claimed-docs] “You can: Connect Claude to existing remote MCP servers. Build your own remote MCP servers to connect with any tool.”
    • [claimed-docs] “You can: - Connect Claude to existing remote MCP servers. - Build your own remote MCP servers to connect with any tool.”
    • [claimed-docs] “installing and managing local MCP servers has become significantly easier... install local MCP servers on your computer as easily as browser…”
    • [claimed-docs] “you can now install local MCP servers on your computer as easily as browser extensions”
    • [claimed-docs] “Our unified directory brings skills, connectors, and plugins together in one place so you can find and install everything that customizes Cl…”
    • [probe] “official MCP server documented at https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp”
    proof ↗⚑ flag

  4. ai-native userUse an official CLI

    weight 2 · round to ChatGPT
    ChatGPTfullprobed7/10

    ChatGPT ships an official Codex CLI with documented commands (--search, --cd, /memories, /fast, MCP config via config.toml, codex-security bulk-scan), and the llms.txt probe confirms 'Codex CLI' as a first-party interface for building and automating dev workflows, satisfying the AI-native official CLI story. Missing for 10: independent/hands-on community verification of the CLI itself, since community evidence only covers the general ChatGPT web product, not Codex CLI usage.

    • [claimed-docs] “In Codex CLI, use /memories in an interactive session to control whether the current chat can use existing local memories or become an input…”
    • [claimed-docs] “Run codex from the directory you want Codex to work in, or pass --cd (-C) to set it explicitly.”
    • [claimed-docs] “In the CLI, pass `--search` to fetch live results for one run”
    • [claimed-docs] “pass `--search` to fetch live results for one run”
    • [claimed-docs] “Use the Codex SDK to automate coding tasks, including jobs in CI.”
    • [claimed-docs] “Inspect code, make changes, run commands, and automate repeatable work without leaving your terminal.”
    • [claimed-docs] “Codex reads AGENTS.md files before doing any work.”
    • [claimed-docs] “Use rules to control which commands Codex can run outside the sandbox.”
    • [claimed-docs] “Use `/fast on`, `/fast off`, or `/fast status` in the CLI to change or inspect the current setting.”
    • [claimed-docs] “Use its command-line interface (CLI) to scan repositories you own or have permission to assess, review findings over time”
    • [claimed-docs] “Use `npx @openai/codex-security bulk-scan` to review repositories in one campaign.”
    • [probe] “PROBE llms.txt: HTTP 200 at https://learn.chatgpt.com/llms.txt # Codex > Build with the Codex CLI, IDE extension, and cloud automation to s…”
    proof ↗⚑ flag
    Claudedisputedcontradicted6/10

    Anthropic ships an official CLI, Claude Code, well documented for terminal-based coding, deployment, debugging, and even screen-reader accessibility (claude-docs-8, claude-docs-23, claude-docs-52, claude-docs-65). However, a hands-on community report describes it as an 'empty unresponsive terminal' upon first use, contradicting the polished experience implied by docs, and another notes serious code-quality issues in the CLI's own codebase (claude-comm-13, claude-comm-15). missing for 10: broader independent corroboration of reliable day-to-day CLI usage, resolution of reported unresponsiveness, and evidence the code-quality issues have been fixed.

    • [claimed-docs] “Build with Claude Code in your terminal, then deploy to a URL Claude can reach. Test and verify in the browser with the Chrome extension. De…”
    • [claimed-docs] “Build with Claude Code in your terminal, then deploy to a URL Claude can reach. Test and verify in the browser with the Chrome extension.”
    • [claimed-docs] “Build with Claude Code in your terminal, then deploy to a URL Claude can reach”
    • [claimed-docs] “It was built with and for screen reader users, and it's useful to anyone who wants plain output for braille displays, slow connections, or t…”
    • [community] “Tried claude code, and have an empty unresponsive terminal. Looks cool in the demo though, but not sure this is going to perform better than…”
    • [community] “src/cli/print.ts is the single worst function in the codebase by every metric: 3,167 lines long, 12 levels of nesting at its deepest, ~486 b…”
    proof ↗⚑ flag

  5. ai-native userDrive the product through a documented public API

    weight 3 · round to ChatGPT
    ChatGPTpartialprobed4/10

    There is documented evidence of programmatic/scriptable access via the Codex CLI and Codex SDK ('Use the Codex SDK to automate coding tasks, including jobs in CI', CLI flags like --search, --cd, /fast, config.toml), which lets an AI-native user drive parts of the product outside the chat UI. However, a probe for a standard documented public API (openapi.json/swagger) at the docs site returned 404 on all candidate paths, and none of the evidence describes a general, versioned public API for driving ChatGPT itself (as opposed to Codex-specific tooling or MCP client configuration). Missing for 10: an explicit REST/GraphQL API reference for ChatGPT product actions, OpenAPI/swagger spec, and independent corroboration that non-Codex ChatGPT features are API-drivable.

    • [claimed-docs] “Use the Codex SDK to automate coding tasks, including jobs in CI.”
    • [claimed-docs] “Inspect code, make changes, run commands, and automate repeatable work without leaving your terminal.”
    • [claimed-docs] “Use its command-line interface (CLI) to scan repositories you own or have permission to assess, review findings over time”
    • [claimed-docs] “Use `npx @openai/codex-security bulk-scan` to review repositories in one campaign.”
    • [probe] “PROBE openapi: all candidate paths 404 (https://learn.chatgpt.com/openapi.json, https://learn.chatgpt.com/swagger.json, https://learn.chatgp…”
    proof ↗⚑ flag
    Claudepartialprobed3/10

    Evidence only hints at API-like surfaces (an Enterprise Compliance API for audit/chat data access, and references to using the Claude API/Console to power products) but never surfaces the actual general-purpose public API documentation for driving Claude's core capabilities; automated probes for an OpenAPI/swagger spec on the docs site returned 404s. Missing for 10: direct citation of the main Claude API reference docs, authentication/quickstart guides, and confirmation of comprehensive public API coverage beyond compliance/audit data.

    • [claimed-docs] “Audit logs: capture key information about user actions, system events, and data access. ... Compliance API: programmatically access Claude u…”
    • [claimed-docs] “Audit logs: capture key information about user actions, system events, and data access.”
    • [claimed-docs] “Enterprise includes everything in the Team plan, plus the following: - Security features to ensure the safety and compliance of your organiz…”
    • [claimed-docs] “You can use the prompt improver in the Claude Console to automatically adapt prompts that were originally written for other AI models.”
    • [probe] “PROBE openapi: all candidate paths 404 (https://support.claude.com/openapi.json, https://support.claude.com/swagger.json, https://support.cl…”
    proof ↗⚑ flag

  6. ai-native userIssue scoped/least-privilege API credentials for an agent

    weight 2 · round to ChatGPT
    ChatGPTpartialclaimed4/10

    ChatGPT/Codex docs show some least-privilege mechanisms for agent tool access — MCP servers can be scoped to a trusted project via `.codex/config.toml` (chatgpt-docs-55), sandbox command rules can restrict what Codex can run outside the sandbox (chatgpt-docs-95), and workspace admins can control access to browser use, plugins, and network access (chatgpt-docs-88), plus GPT builders can configure scoped third-party API actions (chatgpt-docs-34, chatgpt-docs-44). However there is no evidence of a user-facing mechanism to actually issue/generate a scoped or least-privilege API credential/token specifically for an agent to use. missing for 10: explicit scoped API key/token issuance workflow, granularity/expiry controls on such credentials, and any documentation or independent confirmation that a user can mint a restricted-permission credential for agent use.

    • [claimed-docs] “you can also scope MCP servers to a project with `.codex/config.toml` (trusted projects only)”
    • [claimed-docs] “Use rules to control which commands Codex can run outside the sandbox.”
    • [claimed-docs] “Your workspace administrator can control access to ChatGPT Work, plugins, browser use, and network access.”
    • [claimed-docs] “Allow GPT builders to use approved workspace apps or configure actions that interact with permitted third-party APIs.”
    • [claimed-docs] “configure actions that interact with permitted third-party APIs”
    proof ↗⚑ flag
    Claudenone0/10

    No evidence describes scoped or least-privilege API credential issuance for agents; enterprise features mention audit logs and compliance API access but nothing about creating scoped/restricted API keys or credentials for agent use. Missing for 10: any documentation of API key scoping, permission tiers, or least-privilege credential management for agents.

      ⚑ flag

    • ai-native userBuild against official SDKs

      weight 2 · round to ChatGPT
      ChatGPTpartialprobed3/10

      The pack contains a single line noting 'Use the Codex SDK to automate coding tasks, including jobs in CI' (chatgpt-docs-91), indicating an official SDK exists for Codex/automation, but there is no linked documentation, API reference, code samples, or independent corroboration of this SDK's use. The openapi probe found no public API spec, reinforcing that developer-facing SDK documentation is thin in this evidence set. missing for 10: dedicated SDK documentation/reference pages, code examples, language support details, independent developer reports of building with the SDK.

      • [claimed-docs] “Use the Codex SDK to automate coding tasks, including jobs in CI.”
      • [probe] “PROBE openapi: all candidate paths 404 (https://learn.chatgpt.com/openapi.json, https://learn.chatgpt.com/swagger.json, https://learn.chatgp…”
      proof ↗⚑ flag
      Claudenone0/10

      The evidence pack covers Claude's consumer/enterprise features (Cowork, Projects, connectors, MCP, Chrome extension, file handling) but contains no documentation of official language SDKs (e.g., Python/TypeScript client libraries) for building applications against the Claude API. Missing for 10: any first-party SDK docs, API reference, or code samples showing programmatic API usage.

        ⚑ flag

      • ai-native userSubscribe to events via webhooks

        weight 2 · round to ChatGPT
        ChatGPTpartialclaimed4/10

        ChatGPT's Automations/Scheduled Tasks support event-driven triggers ('act on pull request feedback without polling on a fixed cadence') tied to Gmail, Slack, and GitHub events, which functions like a webhook-subscription mechanism, but this is not exposed as a general-purpose webhook subscription API — it's limited to a few pre-integrated services with no documented endpoint/config for arbitrary webhook URLs. Missing for 10: a generic webhook subscription API/config for third-party or custom events, documentation of payload/security handling, and any independent confirmation of reliability.

        • [claimed-docs] “Scheduled tasks can now start when a supported event occurs in Gmail, Slack, or GitHub.”
        • [claimed-docs] “Use an event trigger to triage new email, summarize channel activity, or act on pull request feedback without polling on a fixed cadence.”
        • [claimed-docs] “scheduled tasks can run when a supported Gmail, Slack, or GitHub event occurs”
        proof ↗⚑ flag
        Claudenone0/10

        No evidence pack item describes webhooks or event subscription mechanisms for Claude; the closest agentic integrations are MCP connectors and scheduled tasks, which are not webhook-based event subscriptions.

          ⚑ flag

        Agentic features

        1. ai-native userGet AI-generated insights and suggestions from my data inside the product

          weight 2 · round drawn
          ChatGPTfullclaimed8/10

          ChatGPT can analyze uploaded/connected data (files, projects, local folders, connected apps like Google Drive/SharePoint/Salesforce) and generate insights, summaries, and suggestions (e.g. 'Search the web, browse websites, compare sources, read files, analyze data, and summarize findings' and 'Create a deck, analyze files, draft a report, build a project plan'), with projects giving persistent context for tailored suggestions. Missing for 10: independent hands-on evidence specifically validating data-insight quality/accuracy (community evidence largely covers unrelated features like web search accuracy issues, not data-analysis insight quality).

          • [claimed-docs] “Search the web, browse websites, compare sources, read files, analyze data, and summarize findings.”
          • [claimed-docs] “Create a deck, analyze files, draft a report, build a project plan”
          • [claimed-docs] “Use a project to organize related chats and give ChatGPT the context it needs.”
          • [claimed-docs] “A local project gives chats access to one or more folders on your computer, such as a collection of source files or a codebase.”
          • [claimed-docs] “Plugins can connect ChatGPT to the tools and information you use for work, such as Google Drive, SharePoint, Salesforce, or Gong.”
          • [claimed-docs] “Turn research and analysis into documents, presentations, spreadsheets, and other finished work.”
          proof ↗⚑ flag
          Claudefullclaimed8/10

          Claude can ingest user data (files, connectors to Gmail/Drive/Calendar, Projects knowledge bases) and generate AI-driven insights, analysis, visualizations, and reports directly from that data, including agentic multi-step research that synthesizes findings with citations. This is well documented across file analysis, Research mode, Cowork task automation, and artifact/document generation features. Missing for 10: independent hands-on benchmarking specifically validating insight quality/accuracy from user data (community evidence is mixed/general rather than about this specific data-insight capability).

          • [claimed-docs] “Claude can work with the following document types: - PDF - DOCX - CSV - TXT - HTML - ODT - RTF - EPUB - JSON - XLSX”
          • [claimed-docs] “Prompt Claude using natural language to generate Excel spreadsheets, PowerPoint presentations, Word documents, and PDF files that you can do…”
          • [claimed-docs] “Projects allow you to create self-contained workspaces with their own chat histories and knowledge bases.”
          • [claimed-docs] “With research, Claude delivers thorough answers in minutes, complete with easy-to-check citations so you can trust Claude's findings.”
          • [claimed-docs] “Research transforms how Claude finds and analyzes information. Claude operates agentically, conducting multiple searches that build on each …”
          • [claimed-docs] “Claude operates agentically, conducting multiple searches that build on each other while determining exactly what to investigate next.”
          • [claimed-docs] “produce reports with charts and visualizations, and generate presentations from your documents—all without specialized software skills”
          • [claimed-docs] “Connect your Gmail, Google Calendar, and Google Drive to Claude so you can search and send emails, manage your calendar, work with documents…”
          • [claimed-docs] “Files can be uploaded to individual chats or uploaded to a project's Files section for persistent reference across conversations.”
          proof ↗⚑ flag

        2. ai-native userSet up automations that run autonomously in the background

          weight 2 · round drawn
          ChatGPTfullclaimed8/10

          ChatGPT's Scheduled Tasks/Automations feature explicitly lets users schedule recurring or event-triggered tasks (Gmail, Slack, GitHub events) to run autonomously in the background, with a dashboard to review active, paused, and completed runs. This is well-documented first-party functionality directly matching the story. Missing for 10: independent hands-on verification of background automation reliability and no community corroboration specifically about the automations/scheduled-tasks feature.

          • [claimed-docs] “Schedule recurring tasks to run in the background.”
          • [claimed-docs] “Scheduled tasks can now start when a supported event occurs in Gmail, Slack, or GitHub.”
          • [claimed-docs] “Schedule recurring tasks to run in the background... Review active, paused, and completed tasks and recent runs in Scheduled.”
          • [claimed-docs] “scheduled tasks can run when a supported Gmail, Slack, or GitHub event occurs”
          • [claimed-docs] “Use an event trigger to triage new email, summarize channel activity, or act on pull request feedback without polling on a fixed cadence.”
          proof ↗⚑ flag
          Claudefullclaimed8/10

          Claude Cowork's scheduled tasks let users describe a task once and have Claude execute it autonomously on a recurring or on-demand basis, delivering finished outputs like reports and briefings without further user input, which directly matches background automation. Missing for 10: independent/hands-on verification of scheduled task reliability and no detail on failure handling or notification mechanisms.

          • [claimed-docs] “Instead of starting each task from scratch, you describe it once and Claude handles it on your schedule—delivering finished outputs like rep…”
          • [claimed-docs] “Scheduled tasks allow you to delegate work to Claude Cowork by creating tasks that run automatically on a recurring basis, or on demand.”
          • [claimed-docs] “you describe it once and Claude handles it on your schedule—delivering finished outputs like reports, briefings, and summaries every time”
          • [claimed-docs] “With Cowork, you can describe an outcome, step away, and come back to finished work—formatted documents, organized files, synthesized resear…”
          • [claimed-docs] “Claude can take on complex, multi-step tasks and execute them on your behalf.”
          • [claimed-docs] “brings Claude Code's agentic capabilities to knowledge work beyond coding”
          proof ↗⚑ flag

        3. ai-native userDelegate tasks to a built-in AI assistant inside the product

          weight 3 · round drawn
          ChatGPTfullclaimed9/10

          ChatGPT ships extensive built-in agentic delegation: Work/Codex modes carry tasks through to reviewable results, scheduled/background tasks and event-triggered automations, computer use and browser control, subagent delegation, and voice-initiated task delegation — all first-party features, not third-party add-ons. Missing for 10: independent hands-on verification of these newer agentic features (Work, subagents, Computer Use) beyond vendor docs.

          • [claimed-docs] “Schedule recurring tasks to run in the background.”
          • [claimed-docs] “Scheduled tasks can now start when a supported event occurs in Gmail, Slack, or GitHub.”
          • [claimed-docs] “Turn on Work in the switcher when you want ChatGPT to carry a larger task through to a reviewable result.”
          • [claimed-docs] “Run projects in parallel, work with files, use your computer, and keep long-running work moving from one desktop workspace.”
          • [claimed-docs] “Unlike Work on the Web, local Work can operate on resources that remain on your computer without requiring you to upload files to a cloud co…”
          • [claimed-docs] “In ChatGPT Work, ask ChatGPT to delegate independent work to subagents.”
          • [claimed-docs] “ChatGPT Voice can start separate tasks for longer work, check existing tasks, and send follow-up instructions.”
          • [claimed-docs] “Browser lets ChatGPT open websites, gather current information, and take action while you stay in control.”
          • [claimed-docs] “With Computer Use, ChatGPT can see and operate graphical user interfaces on macOS or Windows.”
          • [claimed-docs] “enter `/goal` to start Goal mode”
          proof ↗⚑ flag
          Claudefullclaimed9/10

          Claude's built-in Cowork feature explicitly lets users delegate multi-step tasks ('describe an outcome, step away, come back to finished work'), with agentic execution including browser/computer use, scheduling, and research that autonomously plans next steps. This is well-documented first-party functionality directly matching task delegation to a built-in assistant. Missing for 10: independent hands-on verification of Cowork's delegation reliability (community evidence covers Claude Code/coding use more than Cowork specifically).

          • [claimed-docs] “With Cowork, you can describe an outcome, step away, and come back to finished work—formatted documents, organized files, synthesized resear…”
          • [claimed-docs] “Instead of starting each task from scratch, you describe it once and Claude handles it on your schedule—delivering finished outputs like rep…”
          • [claimed-docs] “Claude can take on complex, multi-step tasks and execute them on your behalf.”
          • [claimed-docs] “Scheduled tasks allow you to delegate work to Claude Cowork by creating tasks that run automatically on a recurring basis, or on demand.”
          • [claimed-docs] “brings Claude Code's agentic capabilities to knowledge work beyond coding”
          • [claimed-docs] “Research transforms how Claude finds and analyzes information. Claude operates agentically, conducting multiple searches that build on each …”
          • [claimed-docs] “Claude operates agentically, conducting multiple searches that build on each other while determining exactly what to investigate next.”
          proof ↗⚑ flag

        4. ai-native userOperate the product with natural-language commands

          weight 2 · round drawn
          ChatGPTfullcommunity9/10

          ChatGPT is fundamentally a natural-language interface: users type or speak requests and it carries out tasks (browsing, coding, file edits, image generation, scheduling, computer use, voice) via plain conversational commands, per extensive first-party docs and corroborating community reports of natural-language driven capability (e.g. running Docker/filesystem commands via prompts). Missing for 10: independent hands-on verification specifically of newer agentic features (Work, Goal mode, subagents) beyond vendor docs.

          • [claimed-docs] “Browser lets ChatGPT open websites, gather current information, and take action while you stay in control.”
          • [claimed-docs] “With Computer Use, ChatGPT can see and operate graphical user interfaces on macOS or Windows.”
          • [claimed-docs] “Schedule recurring tasks to run in the background.”
          • [claimed-docs] “ChatGPT Voice lets you talk through ideas and coordinate tasks in Chat, Work, and Codex in the ChatGPT desktop app. Start work, check progre…”
          • [claimed-docs] “In the ChatGPT desktop app, enter `/goal` to start Goal mode.”
          • [claimed-docs] “In ChatGPT Work, ask ChatGPT to delegate independent work to subagents.”
          • [community] “I am beyond astounded. I was able to run a Docker image, utilize the fs inside of the container, and exit the container. Docker system comma…”
          • [community] “This is huge, essentially adding what people have been building with LangChain Tools into the core product. The browser and file-upload/inte…”
          proof ↗⚑ flag
          Claudefullcommunity9/10

          Claude's entire interface is natural-language chat/voice, and documentation shows this extends to agentic actions—generating files, browsing/clicking the web, using the computer, scheduling recurring tasks, and building artifacts—all triggered by describing the desired outcome in plain language (claude-docs-4,6,7,15,37,38,54). Community evidence corroborates real-world agentic use (claude-comm-1,10,12), though some report friction with CLI usability. Missing for 10: independent benchmark of NL command reliability across all surfaces, and no rebuttal to the CLI unresponsiveness anecdote being addressed.

          • [claimed-docs] “With Cowork, you can describe an outcome, step away, and come back to finished work—formatted documents, organized files, synthesized resear…”
          • [claimed-docs] “Claude opens sites, reads pages, clicks, types, and fills forms while you watch, with no need to switch windows.”
          • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would.”
          • [claimed-docs] “Prompt Claude using natural language to generate Excel spreadsheets, PowerPoint presentations, Word documents, and PDF files that you can do…”
          • [claimed-docs] “Claude can take on complex, multi-step tasks and execute them on your behalf.”
          • [claimed-docs] “Scheduled tasks allow you to delegate work to Claude Cowork by creating tasks that run automatically on a recurring basis, or on demand.”
          • [claimed-docs] “build tools, visualizations, and experiences by simply describing what you need”
          • [community] “I've been running Opus 4.8 for agentic coding and I don't see it being significantly better than Sonnet 4.5. I find that pairing Google Gemi…”
          • [community] “Anecdotal, but it 1 shot fixed a UI bug that neither Opus 4.5/Codex 5.2-high could fix.”
          • [community] “Claude is significantly better than other models at code assistant tasks, or at least in the way I use it.”
          proof ↗⚑ flag

        Api quality

        1. ai-native userExplore an interactive API reference with runnable examples

          weight 2 · round drawn
          ChatGPTnone0/10

          The evidence pack contains no documentation of an interactive API reference or runnable code examples; a direct probe for OpenAPI/swagger specs on the docs site returned 404 across all candidate paths, and no other citation mentions an API reference sandbox or runnable snippets.

          • [probe] “PROBE openapi: all candidate paths 404 (https://learn.chatgpt.com/openapi.json, https://learn.chatgpt.com/swagger.json, https://learn.chatgp…”
          proof ↗⚑ flag
          Claudenone0/10

          This evidence pack is entirely about Claude's consumer/product features (Cowork, connectors, memory, artifacts, etc.); there is no evidence of an interactive API reference with runnable examples. The probe explicitly found no OpenAPI spec at the checked endpoints, and no docs page describing an interactive API playground is present.

          • [probe] “PROBE openapi: all candidate paths 404 (https://support.claude.com/openapi.json, https://support.claude.com/swagger.json, https://support.cl…”
          • [probe] “PROBE docs-md: HTTP 404 at https://support.claude.com/en/.md”
          proof ↗⚑ flag

        2. ai-native userDownload a machine-readable API spec (OpenAPI or equivalent)

          weight 2 · round drawn
          ChatGPTnone0/10

          The probe explicitly checked for an OpenAPI/machine-readable spec at all standard locations and found only 404s, with no other evidence pack item showing a downloadable API spec for ChatGPT itself.

          • [probe] “PROBE openapi: all candidate paths 404 (https://learn.chatgpt.com/openapi.json, https://learn.chatgpt.com/swagger.json, https://learn.chatgp…”
          proof ↗⚑ flag
          Claudenone0/10

          The evidence pack includes explicit probes checking for a machine-readable API spec (openapi.json, swagger.json, etc.) on Claude's support domain, all returning 404, and no docs item anywhere references an OpenAPI/Swagger spec or downloadable schema for the Claude API. Nothing in the docs list an API reference format for AI-native consumption.

          • [probe] “PROBE openapi: all candidate paths 404 (https://support.claude.com/openapi.json, https://support.claude.com/swagger.json, https://support.cl…”
          • [probe] “PROBE docs-md: HTTP 404 at https://support.claude.com/en/.md”
          proof ↗⚑ flag

        3. ai-native userTest against a sandbox environment without touching production data

          weight 1 · round to ChatGPT
          ChatGPTpartialclaimed4/10

          Codex documentation indicates commands run inside a sandbox by default, with rules governing what can run 'outside the sandbox' (chatgpt-docs-95), and Codex Security's scanning is scoped to repos you own/have permission to assess (chatgpt-docs-97/98), implying isolated execution rather than direct production access. However, there is no explicit documentation describing a dedicated 'test/staging' environment distinct from production data, no detail on how production systems/data are excluded, and no independent/hands-on confirmation of this isolation. Missing for 10: explicit sandbox-vs-production data separation docs, details on network/data isolation guarantees, and independent verification of the sandbox boundary holding in practice.

          • [claimed-docs] “Use rules to control which commands Codex can run outside the sandbox.”
          • [claimed-docs] “Use its command-line interface (CLI) to scan repositories you own or have permission to assess, review findings over time”
          • [claimed-docs] “Use `npx @openai/codex-security bulk-scan` to review repositories in one campaign.”
          proof ↗⚑ flag
          Claudenone0/10

          The evidence pack shows Claude's connectors, Cowork, computer use, and browser automation acting directly on real accounts (Gmail, Drive, live websites) with no mention of a sandbox, staging, or test-mode environment that isolates actions from production data.

            ⚑ flag

          • ai-native userRely on versioned APIs with a documented deprecation policy

            weight 2 · round drawn
            ChatGPTnone0/10

            No evidence of a versioned API or a documented deprecation policy is present; the OpenAPI probe explicitly returned 404s and none of the docs mention API versioning or deprecation practices.

            • [probe] “PROBE openapi: all candidate paths 404 (https://learn.chatgpt.com/openapi.json, https://learn.chatgpt.com/swagger.json, https://learn.chatgp…”
            proof ↗⚑ flag
            Claudenone0/10

            The evidence pack covers Claude's consumer features, connectors, Cowork, and MCP integrations, but contains no mention of API versioning schemes or a documented deprecation policy for any Claude/Anthropic API. Since Claude does expose developer-facing APIs (e.g., for Console, Claude Code, MCP), this axis is applicable, but no evidence supports it.

              ⚑ flag

            Agents tasks — stories about agents tasks in this arenaAgents tasks

            Stories about agents tasks in this arena

            Agent mode

            1. power-userDelegate a multi-step task that the assistant works on autonomously in the background and returns for my review

              weight 3 · round drawn
              ChatGPTfullclaimed8/10

              ChatGPT's Tasks/automations feature explicitly supports scheduling and running multi-step work in the background (including event-triggered runs from Gmail/Slack/GitHub), with a review interface for active/completed runs, and 'Work' mode is described as carrying larger tasks through to a reviewable result; Codex further supports background coding tasks and subagent delegation. Missing for 10: independent hands-on verification of background task reliability/quality and no third-party review of the review/approval workflow.

              • [claimed-docs] “Schedule recurring tasks to run in the background.”
              • [claimed-docs] “Scheduled tasks can now start when a supported event occurs in Gmail, Slack, or GitHub.”
              • [claimed-docs] “Schedule recurring tasks to run in the background... Review active, paused, and completed tasks and recent runs in Scheduled.”
              • [claimed-docs] “Turn on Work in the switcher when you want ChatGPT to carry a larger task through to a reviewable result.”
              • [claimed-docs] “scheduled tasks can run when a supported Gmail, Slack, or GitHub event occurs”
              • [claimed-docs] “In ChatGPT Work, ask ChatGPT to delegate independent work to subagents.”
              • [claimed-docs] “Start a Codex task to run the tests and investigate anything that doesn't pass.”
              proof ↗⚑ flag
              Claudefullclaimed8/10

              Claude Cowork is documented as letting users describe a multi-step outcome, step away, and return to finished work, with scheduled/recurring tasks and background agentic execution (browser/computer use, file generation) explicitly designed for delegation and later review. Missing for 10: independent hands-on validation of Cowork's reliability/quality for complex delegated tasks and clearer detail on review/approval workflow beyond docs.

              • [claimed-docs] “With Cowork, you can describe an outcome, step away, and come back to finished work—formatted documents, organized files, synthesized resear…”
              • [claimed-docs] “Instead of starting each task from scratch, you describe it once and Claude handles it on your schedule—delivering finished outputs like rep…”
              • [claimed-docs] “Claude can take on complex, multi-step tasks and execute them on your behalf.”
              • [claimed-docs] “Scheduled tasks allow you to delegate work to Claude Cowork by creating tasks that run automatically on a recurring basis, or on demand.”
              • [claimed-docs] “you describe it once and Claude handles it on your schedule—delivering finished outputs like reports, briefings, and summaries every time”
              • [claimed-docs] “This article explains how to use Claude Cowork, which brings Claude Code's agentic capabilities to knowledge work beyond coding.”
              • [claimed-docs] “Claude opens sites, reads pages, clicks, types, and fills forms while you watch, with no need to switch windows.”
              • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would.”
              proof ↗⚑ flag

            2. power-userHave the assistant operate a web browser on my behalf to research and complete tasks on websites

              weight 3 · round drawn
              ChatGPTfullclaimed8/10

              ChatGPT's Browser feature explicitly lets it open websites, gather information, and take action on the user's behalf while staying in control, and can complete multi-step tasks like comparing options or completing actions on a website; this is reinforced by Computer Use (operating GUIs) and Work/Codex browser tab integration for signed-in site tasks. missing for 10: independent hands-on verification of complex multi-step website task completion, and detail on reliability/success rate limits.

              • [claimed-docs] “Browser lets ChatGPT open websites, gather current information, and take action while you stay in control.”
              • [claimed-docs] “Use it to compare options, complete a multi-step task on a website, or review a page you're building.”
              • [claimed-docs] “With Computer Use, ChatGPT can see and operate graphical user interfaces on macOS or Windows.”
              • [claimed-docs] “ChatGPT can see and operate graphical user interfaces on macOS or Windows”
              • [claimed-docs] “Bring an open tab into a ChatGPT Work or Codex chat and work with the website where you're already signed in.”
              • [claimed-docs] “Follow the sign-in request and enter your details in the sign-in flow, not in the chat. This doesn't connect your local browser profile.”
              • [claimed-docs] “With Site tools (WebMCP), ChatGPT Work and Codex can use actions offered by a website in the desktop app's built-in browser.”
              proof ↗⚑ flag
              Claudefullclaimed8/10

              Claude has multiple documented browser-control capabilities: the built-in browser in Cowork that opens sites, reads pages, clicks, types, and fills forms autonomously (claude-docs-6/67), the Claude in Chrome extension that reads/clicks/navigates websites (claude-docs-63), and computer-use navigation for on-screen actions (claude-docs-7/22/59), enabling research and task completion on websites. Missing for 10: independent hands-on validation of browser-task success rates and reliability under real-world site complexity.

              • [claimed-docs] “Claude opens sites, reads pages, clicks, types, and fills forms while you watch, with no need to switch windows.”
              • [claimed-docs] “Claude opens sites, reads pages, clicks, types, and fills forms while you watch, with no need to swit”
              • [claimed-docs] “Claude in Chrome is a browser extension that allows Claude to read, click, and navigate websites alongside you.”
              • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would.”
              • [claimed-docs] “When computer use is enabled and Claude doesn't have a connector or tool for what you need, it may navigate to your screen directly—clicking…”
              • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would”
              • [claimed-docs] “Claude can take on complex, multi-step tasks and execute them on your behalf.”
              proof ↗⚑ flag

            3. power-userLet the assistant see and operate applications on my computer to complete work

              weight 2 · round to ChatGPT
              ChatGPTfullclaimed9/10

              ChatGPT's Computer Use feature explicitly lets it see and operate GUIs on macOS/Windows, including testing desktop/mobile app flows, backed by Browser and Work agent capabilities that carry out multi-step tasks on the user's behalf. Documentation is extensive and detailed across multiple first-party sources describing screen operation, browser control, and file/app interaction. Missing for 10: independent hands-on verification of Computer Use reliability in practice.

              • [claimed-docs] “With Computer Use, ChatGPT can see and operate graphical user interfaces on macOS or Windows.”
              • [claimed-docs] “ChatGPT can see and operate graphical user interfaces on macOS or Windows”
              • [claimed-docs] “ChatGPT can see and operate graphical user interfaces on macOS or Windows.”
              • [claimed-docs] “Testing a macOS app, Windows app, iOS simulator flow, or another desktop app that ChatGPT is building.”
              • [claimed-docs] “Browser lets ChatGPT open websites, gather current information, and take action while you stay in control.”
              • [claimed-docs] “Use it to compare options, complete a multi-step task on a website, or review a page you're building.”
              • [claimed-docs] “Run projects in parallel, work with files, use your computer, and keep long-running work moving from one desktop workspace.”
              • [claimed-docs] “Unlike Work on the Web, local Work can operate on resources that remain on your computer without requiring you to upload files to a cloud co…”
              • [claimed-docs] “Your ChatGPT data controls apply to content processed through ChatGPT, including screenshots taken by Computer Use.”
              proof ↗⚑ flag
              Claudefullclaimed8/10

              Claude Cowork/computer use lets Claude navigate directly to the user's screen—clicking, typing, opening apps, filling forms—while the user watches, plus a built-in browser for opening sites and interacting with pages, directly matching the story of operating applications to complete work. Community evidence corroborates agentic coding/task use though notes mixed quality perceptions unrelated to this specific capability. Missing for 10: independent hands-on verification of computer-use reliability/accuracy and broader third-party benchmarking beyond vendor docs.

              • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would.”
              • [claimed-docs] “When computer use is enabled and Claude doesn't have a connector or tool for what you need, it may navigate to your screen directly—clicking…”
              • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would”
              • [claimed-docs] “Claude opens sites, reads pages, clicks, types, and fills forms while you watch, with no need to switch windows.”
              • [claimed-docs] “Claude opens sites, reads pages, clicks, types, and fills forms while you watch, with no need to swit”
              • [claimed-docs] “With Cowork, you can describe an outcome, step away, and come back to finished work—formatted documents, organized files, synthesized resear…”
              • [claimed-docs] “Claude can take on complex, multi-step tasks and execute them on your behalf.”
              • [claimed-docs] “Claude in Chrome is a browser extension that allows Claude to read, click, and navigate websites alongside you.”
              proof ↗⚑ flag

            Tasks

            1. power-userSchedule recurring or one-off tasks that run automatically and come back to me with results

              weight 2 · round drawn
              ChatGPTfullclaimed8/10

              ChatGPT's Automations/Scheduled Tasks feature explicitly supports scheduling recurring or one-off tasks that run in the background and report back, plus event-triggered tasks from Gmail/Slack/GitHub, with a UI to review active, paused, and completed runs. missing for 10: independent/hands-on community corroboration of scheduled task reliability and no detail on notification/result-delivery mechanics beyond docs.

              • [claimed-docs] “Schedule recurring tasks to run in the background.”
              • [claimed-docs] “Scheduled tasks can now start when a supported event occurs in Gmail, Slack, or GitHub.”
              • [claimed-docs] “Schedule recurring tasks to run in the background... Review active, paused, and completed tasks and recent runs in Scheduled.”
              • [claimed-docs] “scheduled tasks can run when a supported Gmail, Slack, or GitHub event occurs”
              • [claimed-docs] “Use an event trigger to triage new email, summarize channel activity, or act on pull request feedback without polling on a fixed cadence.”
              proof ↗⚑ flag
              Claudefullclaimed8/10

              Claude Cowork explicitly supports scheduled tasks that run on a recurring or one-off basis and deliver finished outputs like reports and summaries back to the user, matching the story closely. Missing for 10: independent/hands-on validation of scheduling reliability and no detail on notification/delivery mechanisms beyond docs.

              • [claimed-docs] “Scheduled tasks allow you to delegate work to Claude Cowork by creating tasks that run automatically on a recurring basis, or on demand.”
              • [claimed-docs] “you describe it once and Claude handles it on your schedule—delivering finished outputs like reports, briefings, and summaries every time”
              • [claimed-docs] “Instead of starting each task from scratch, you describe it once and Claude handles it on your schedule—delivering finished outputs like rep…”
              • [claimed-docs] “With Cowork, you can describe an outcome, step away, and come back to finished work—formatted documents, organized files, synthesized resear…”
              • [claimed-docs] “Claude can take on complex, multi-step tasks and execute them on your behalf.”
              proof ↗⚑ flag

            Apps devices — stories about apps devices in this arenaApps devices

            Stories about apps devices in this arena

            Apps

            1. power-userUse an official desktop app with OS-level shortcuts and access to what is on my screen

              weight 2 · round to ChatGPT
              ChatGPTpartialclaimed6/10

              The evidence confirms an official ChatGPT desktop app (macOS/Windows) that can 'use your computer' and run 'Computer Use' to see and operate GUIs, including screenshot-based screen access, plus desktop-only features like floating Pets controls and multi-browser support. However, there is no documentation of OS-level global keyboard shortcuts (e.g., a system-wide hotkey to invoke the app) or of general 'what's on my screen' querying outside the specific Computer Use/testing use case. Missing for 10: explicit OS-level shortcut/hotkey documentation, general screen-content awareness beyond Computer Use testing scenarios, independent hands-on confirmation of these desktop-specific features.

              • [claimed-docs] “Run projects in parallel, work with files, use your computer, and keep long-running work moving from one desktop workspace.”
              • [claimed-docs] “Work in Edge, Brave, Opera, or Vivaldi as well as Chrome from the ChatGPT desktop app.”
              • [claimed-docs] “With Computer Use, ChatGPT can see and operate graphical user interfaces on macOS or Windows.”
              • [claimed-docs] “ChatGPT can see and operate graphical user interfaces on macOS or Windows”
              • [claimed-docs] “ChatGPT can see and operate graphical user interfaces on macOS or Windows.”
              • [claimed-docs] “Your ChatGPT data controls apply to content processed through ChatGPT, including screenshots taken by Computer Use.”
              • [claimed-docs] “Type a request or start a voice conversation from the floating Pets controls in the ChatGPT desktop app on macOS and Windows.”
              • [claimed-docs] “Testing a macOS app, Windows app, iOS simulator flow, or another desktop app that ChatGPT is building.”
              proof ↗⚑ flag
              Claudepartialclaimed5/10

              Claude ships an official desktop app (Mac, Windows, Linux via apt) and a Cowork/computer-use feature that lets Claude 'navigate to your screen directly—clicking, typing, and opening apps' while the user watches, which shows real screen access. However there is no evidence of dedicated OS-level keyboard shortcuts (e.g., a global hotkey to invoke Claude with current screen context) — the screen access described is agentic task automation rather than a power-user shortcut workflow. Missing for 10: documented OS-level global shortcuts/hotkeys, and evidence that a user can quickly summon Claude to see the current screen via keypress rather than launching a Cowork/computer-use task.

              • [claimed-docs] “you can install Claude Desktop from Anthropic's apt repository rather than as a downloaded .deb file so that updates arrive through your sys…”
              • [claimed-docs] “The Claude desktop apps bring Claude's capabilities directly to your computer, allowing for seamless integration with your workflow.”
              • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would.”
              • [claimed-docs] “When computer use is enabled and Claude doesn't have a connector or tool for what you need, it may navigate to your screen directly—clicking…”
              • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would”
              • [claimed-docs] “This article explains how to use Claude Cowork, which brings Claude Code's agentic capabilities to knowledge work beyond coding.”
              proof ↗⚑ flag

            2. knowledge-workerUse full-featured official mobile apps for iOS and Android

              weight 2 · round to Claude
              ChatGPTfullclaimed7/10

              The iOS App Store listing (chatgpt-docs-20/21/41/52/58/59) documents rich mobile features—voice mode, photo upload, image generation, stickers—confirming a full-featured iOS app, and doc references (chatgpt-docs-67) explicitly mention continuing tasks 'in ChatGPT Work on the web, iOS, or Android,' indicating Android parity. Missing for 10: no dedicated Android app store evidence pack, no independent hands-on reviews of the mobile apps' feature completeness or performance.

              • [claimed-docs] “Photo upload—Snap or upload a picture to transcribe a handwritten recipe or get info about a landmark.”
              • [claimed-docs] “Advanced Voice Mode–Tap the soundwave icon to have a real-time convo on the go.”
              • [claimed-docs] “Advanced Voice Mode–Tap the soundwave icon to have a real-time convo on the go. Settle a dinner table debate, or practice a new language.”
              • [claimed-docs] “Snap or upload a picture to transcribe a handwritten recipe or get info about a landmark.”
              • [claimed-docs] “Image generation–Generate original images from a description, or transform existing ones with a few simple words.”
              • [claimed-docs] “Choose an image and style, remix with emojis, and make personalized stickers that are ready to share in your group chat.”
              • [claimed-docs] “continue a task that needs a website account in ChatGPT Work on the web, iOS, or Android”
              proof ↗⚑ flag
              Claudefullclaimed8/10

              Official iOS and Android apps are documented (App Store install instructions, Chrome/mobile chat parity across web/iOS/Android/desktop, voice mode explicitly available on Claude Mobile), indicating full-featured mobile apps rather than a bare wrapper. Missing for 10: no independent hands-on review of mobile app feature parity/quality, and no detail on which advanced features (Cowork, computer use) are available on mobile vs desktop-only.

              • [claimed-docs] “You can install the Claude app onto your iOS device by navigating to the App Store and searching for “Claude by Anthropic””
              • [claimed-docs] “You can install the Claude app onto your iOS device by navigating to the App Store and searching for "Claude by Anthropic"”
              • [claimed-docs] “Chat on web, iOS, Android, and on your desktop ... Memory across conversations”
              • [claimed-docs] “Chat on web, iOS, Android, and on your desktop”
              • [claimed-docs] “Voice mode is a beta feature available to all plans (Free, Pro, Max, Team, and Enterprise) on Claude Mobile (iOS and Android), Claude Deskto…”
              proof ↗⚑ flag

            Custom bots

            1. power-userBuild and share custom assistants with their own instructions and knowledge

              weight 2 · round to Claude
              ChatGPTpartialclaimed5/10

              Docs confirm a 'GPT builder' concept for creating custom GPTs with configurable actions/connected apps and a dedicated 'gpts-and-sharing' doc implying sharing, but no evidence details custom instructions or uploading a knowledge base, and no independent/hands-on corroboration of the sharing flow. Missing for 10: explicit mention of setting custom instructions, uploading knowledge files, and community/hands-on validation of building and sharing a custom GPT.

              • [claimed-docs] “Allow GPT builders to use approved workspace apps or configure actions that interact with permitted third-party APIs.”
              • [claimed-docs] “configure actions that interact with permitted third-party APIs”
              • [claimed-docs] “GPT builders can use either connected apps or custom actions, but not both in the same GPT.”
              proof ↗⚑ flag
              Claudepartialclaimed6/10

              Claude supports Projects (self-contained workspaces with custom knowledge bases/files) and Skills (teach Claude repeatable instructions like brand guidelines), which together let a power-user build a persona-like assistant with instructions and knowledge; Projects can be shared with team members. However, there's no dedicated 'custom GPT'-style public sharing/marketplace for assistants, and no independent evidence of end-to-end sharing outside an org. missing for 10: evidence of public/marketplace sharing of custom assistants, explicit persona/system-instruction configuration UI, and independent hands-on confirmation of building and sharing such assistants.

              • [claimed-docs] “Projects allow you to create self-contained workspaces with their own chat histories and knowledge bases.”
              • [claimed-docs] “whether that's creating documents with your company's brand guidelines, analyzing data using your organization's specific workflows”
              • [claimed-docs] “Skills teach Claude how to complete specific tasks in a repeatable way, whether that's creating documents with your company's brand guidelin…”
              • [claimed-docs] “Files can be uploaded to individual chats or uploaded to a project's Files section for persistent reference across conversations.”
              proof ↗⚑ flag

            Automation depth — how much of the product can run unattendedAutomation depth

            How much of the product can run unattended

            1. ai-native userPerform bulk operations across many items at once

              weight 2 · round to ChatGPT
              ChatGPTpartialclaimed6/10

              ChatGPT/Codex docs show concrete bulk-style automation: 'bulk-scan' scanning many repositories in one campaign (chatgpt-docs-98), a CLI for scanning multiple repos over time (chatgpt-docs-97), running multiple projects in parallel (chatgpt-docs-32), and delegating work to subagents (chatgpt-docs-103), plus background scheduled tasks with multiple runs (chatgpt-docs-37). This covers meaningful bulk/parallel automation but is concentrated in coding/security contexts rather than general-purpose bulk operations across arbitrary item sets, and there's no independent/hands-on corroboration of these bulk claims. Missing for 10: evidence of bulk operations on non-code items (e.g., bulk document/data processing), and independent verification that bulk-scan/subagent delegation works reliably at scale.

              • [claimed-docs] “Use `npx @openai/codex-security bulk-scan` to review repositories in one campaign.”
              • [claimed-docs] “Use its command-line interface (CLI) to scan repositories you own or have permission to assess, review findings over time”
              • [claimed-docs] “Run projects in parallel, work with files, use your computer, and keep long-running work moving from one desktop workspace.”
              • [claimed-docs] “In ChatGPT Work, ask ChatGPT to delegate independent work to subagents.”
              • [claimed-docs] “Schedule recurring tasks to run in the background... Review active, paused, and completed tasks and recent runs in Scheduled.”
              proof ↗⚑ flag
              Claudenone0/10

              The evidence pack documents agentic task automation (Cowork, scheduled tasks, Research) and single-document file handling, but no feature is described for processing or acting on many items at once (e.g., batch file processing, bulk edit across records, or a batch API). Automation-depth stories like this are plausible for Claude given its agentic tooling, but nothing in the pack shows bulk/multi-item operation support.

                ⚑ flag

              • ai-native userDefine rules that trigger actions automatically on events

                weight 3 · round to ChatGPT
                ChatGPTfullclaimed7/10

                ChatGPT's automations/scheduled tasks feature explicitly supports event-driven triggers (Gmail, Slack, GitHub) instead of just fixed schedules, letting users define rules that fire actions on external events, and lets them review active/paused/completed runs. Missing for 10: independent or hands-on verification that these event triggers work reliably in practice, and detail on how flexible/general the rule definitions can be beyond the three named integrations.

                • [claimed-docs] “Schedule recurring tasks to run in the background.”
                • [claimed-docs] “Scheduled tasks can now start when a supported event occurs in Gmail, Slack, or GitHub.”
                • [claimed-docs] “Use an event trigger to triage new email, summarize channel activity, or act on pull request feedback without polling on a fixed cadence.”
                • [claimed-docs] “Schedule recurring tasks to run in the background... Review active, paused, and completed tasks and recent runs in Scheduled.”
                • [claimed-docs] “scheduled tasks can run when a supported Gmail, Slack, or GitHub event occurs”
                proof ↗⚑ flag
                Claudepartialclaimed4/10

                Claude Cowork supports scheduled/recurring tasks that run automatically on a time-based schedule or on demand (claude-docs-38, claude-docs-58), which is a form of automation, but there's no evidence of user-defined rules triggered by external events (e.g., 'when an email arrives' or 'when a file changes, do X') as opposed to calendar/time-based scheduling. missing for 10: event-driven trigger conditions (webhooks, connector-based event listeners), a rules engine for conditional automation, and any documentation of non-time-based triggers.

                • [claimed-docs] “Scheduled tasks allow you to delegate work to Claude Cowork by creating tasks that run automatically on a recurring basis, or on demand.”
                • [claimed-docs] “you describe it once and Claude handles it on your schedule—delivering finished outputs like reports, briefings, and summaries every time”
                • [claimed-docs] “Instead of starting each task from scratch, you describe it once and Claude handles it on your schedule—delivering finished outputs like rep…”
                • [claimed-docs] “Claude can take on complex, multi-step tasks and execute them on your behalf.”
                proof ↗⚑ flag

              • ai-native userSchedule recurring jobs or workflows

                weight 2 · round drawn
                ChatGPTfullclaimed8/10

                ChatGPT's Scheduled Tasks/Automations feature explicitly supports recurring background jobs ('Schedule recurring tasks to run in the background,' with review of active/paused/completed runs) and even event-based triggers (Gmail, Slack, GitHub) as an alternative to fixed cadence. This directly matches the story of scheduling recurring jobs/workflows with first-party docs. Missing for 10: independent/hands-on community corroboration of the scheduling feature specifically (evidence pack community items don't address automations/scheduling) and more detail on reliability/limits in practice.

                • [claimed-docs] “Schedule recurring tasks to run in the background.”
                • [claimed-docs] “Schedule recurring tasks to run in the background... Review active, paused, and completed tasks and recent runs in Scheduled.”
                • [claimed-docs] “Scheduled tasks can now start when a supported event occurs in Gmail, Slack, or GitHub.”
                • [claimed-docs] “scheduled tasks can run when a supported Gmail, Slack, or GitHub event occurs”
                • [claimed-docs] “Use an event trigger to triage new email, summarize channel activity, or act on pull request feedback without polling on a fixed cadence.”
                proof ↗⚑ flag
                Claudefullclaimed8/10

                Claude Cowork explicitly supports scheduled recurring tasks, letting users describe a workflow once and have Claude execute it automatically on a recurring or on-demand basis, delivering outputs like reports and briefings — this directly matches the story. Missing for 10: independent/hands-on verification of scheduling reliability and details on scheduling granularity/limits beyond vendor docs.

                • [claimed-docs] “Instead of starting each task from scratch, you describe it once and Claude handles it on your schedule—delivering finished outputs like rep…”
                • [claimed-docs] “Scheduled tasks allow you to delegate work to Claude Cowork by creating tasks that run automatically on a recurring basis, or on demand.”
                • [claimed-docs] “you describe it once and Claude handles it on your schedule—delivering finished outputs like reports, briefings, and summaries every time”
                • [claimed-docs] “With Cowork, you can describe an outcome, step away, and come back to finished work—formatted documents, organized files, synthesized resear…”
                proof ↗⚑ flag

              • ai-native userVersion, review, and roll back my automations

                weight 1 · round to ChatGPT
                ChatGPTpartialclaimed4/10

                ChatGPT supports scheduled/recurring automations with a review of active/paused/completed task runs (chatgpt-docs-11, chatgpt-docs-37, chatgpt-docs-47), which gives some review and status visibility, but there is no evidence of version history, diffing between automation versions, or a rollback mechanism to revert an automation to a prior state. missing for 10: explicit versioning of automations, diff/comparison between versions, and a documented rollback/revert feature.

                • [claimed-docs] “Schedule recurring tasks to run in the background.”
                • [claimed-docs] “Schedule recurring tasks to run in the background... Review active, paused, and completed tasks and recent runs in Scheduled.”
                • [claimed-docs] “scheduled tasks can run when a supported Gmail, Slack, or GitHub event occurs”
                proof ↗⚑ flag
                Claudenone0/10

                No evidence describes versioning, reviewing, or rolling back automations like scheduled Cowork tasks or Skills; docs cover creating/scheduling tasks but not history, diffs, or rollback capability.

                  ⚑ flag

                Connectors apps — stories about connectors apps in this arenaConnectors apps

                Stories about connectors apps in this arena

                Connectors

                1. power-userBrowse a directory of third-party apps and connectors and add them to the assistant

                  weight 2 · round to Claude
                  ChatGPTfullclaimed7/10

                  ChatGPT documents a Plugins/connectors ecosystem — installing named plugins (Slack, Codex Security), connecting to third-party tools like Google Drive, SharePoint, Salesforce, and Gong, and GPT builders choosing 'connected apps' or custom actions from an approved list — which together describe browsing and adding third-party integrations to the assistant. Missing for 10: independent/hands-on confirmation of the actual browse-and-add directory UI/UX and its breadth beyond the named examples.

                  • [claimed-docs] “Install the Slack plugin to summarize channels or draft replies.”
                  • [claimed-docs] “Install the Codex Security plugin to scan authorized code and confirm plausible vulnerability findings.”
                  • [claimed-docs] “Plugins can connect ChatGPT to the tools and information you use for work, such as Google Drive, SharePoint, Salesforce, or Gong.”
                  • [claimed-docs] “Allow GPT builders to use approved workspace apps or configure actions that interact with permitted third-party APIs.”
                  • [claimed-docs] “GPT builders can use either connected apps or custom actions, but not both in the same GPT.”
                  • [claimed-docs] “configure actions that interact with permitted third-party APIs”
                  proof ↗⚑ flag
                  Claudefullclaimed9/10

                  Claude documents a unified directory that brings skills, connectors, and plugins together in one place to find and install everything that customizes Claude, plus specific connectors like Google Workspace and remote/local MCP servers to extend the assistant. Missing for 10: no independent hands-on review confirming the browsing/discovery UX of the directory itself.

                  • [claimed-docs] “Our unified directory brings skills, connectors, and plugins together in one place so you can find and install everything that customizes Cl…”
                  • [claimed-docs] “Our unified directory brings skills, connectors, and plugins together in one place so you can find and install everything that customizes Cl…”
                  • [claimed-docs] “Connect your Gmail, Google Calendar, and Google Drive to Claude so you can search and send emails, manage your calendar, work with documents…”
                  • [claimed-docs] “Connect your Gmail, Google Calendar, and Google Drive to Claude so you can search and send emails, manage your calendar, work with documents…”
                  • [claimed-docs] “installing and managing local MCP servers has become significantly easier... install local MCP servers on your computer as easily as browser…”
                  • [claimed-docs] “you can now install local MCP servers on your computer as easily as browser extensions”
                  • [claimed-docs] “You can: Connect Claude to existing remote MCP servers. Build your own remote MCP servers to connect with any tool.”
                  proof ↗⚑ flag

                2. knowledge-workerConnect my cloud drive, email, and calendar so the assistant can search and use them in answers

                  weight 3 · round to Claude
                  ChatGPTpartialclaimed6/10

                  Docs show ChatGPT can connect to Google Drive/SharePoint via plugins and can trigger on Gmail events, giving cloud-drive and email connectivity, and MCP/plugins provide a general connector framework for third-party tools. However, there is no explicit mention of calendar integration, and all evidence is vendor documentation with no independent/hands-on confirmation that these connectors actually surface content in answers. Missing for 10: explicit calendar connector support, independent corroboration of connector functionality, and clearer 'search across all three data sources in one answer' evidence.

                  • [claimed-docs] “Plugins can connect ChatGPT to the tools and information you use for work, such as Google Drive, SharePoint, Salesforce, or Gong.”
                  • [claimed-docs] “Install the Slack plugin to summarize channels or draft replies.”
                  • [claimed-docs] “Scheduled tasks can now start when a supported event occurs in Gmail, Slack, or GitHub.”
                  • [claimed-docs] “scheduled tasks can run when a supported Gmail, Slack, or GitHub event occurs”
                  • [claimed-docs] “Model Context Protocol (MCP) connects models to tools and context. Use it to give ChatGPT or Codex access to third-party documentation”
                  • [claimed-docs] “Use it to give ChatGPT or Codex access to third-party documentation, or to let it interact with developer tools like your browser or Figma.”
                  proof ↗⚑ flag
                  Claudefullclaimed9/10

                  Claude has a documented Google Workspace connector letting users connect Gmail, Google Calendar, and Google Drive so Claude can search emails, manage calendar, and work with documents/files directly in conversation, plus a privacy commitment not to train on this data. This directly satisfies the story's cloud drive/email/calendar connection use case. Missing for 10: no independent hands-on report corroborating real-world reliability of these specific connectors.

                  • [claimed-docs] “Connect your Gmail, Google Calendar, and Google Drive to Claude so you can search and send emails, manage your calendar, work with documents…”
                  • [claimed-docs] “Connect your Gmail, Google Calendar, and Google Drive to Claude so you can search and send emails, manage your calendar, work with documents…”
                  • [claimed-docs] “We do not train our models on your Gmail, Drive, or Calendar connector data, ensuring your private information remains private.”
                  • [claimed-docs] “Our unified directory brings skills, connectors, and plugins together in one place so you can find and install everything that customizes Cl…”
                  proof ↗⚑ flag

                Files analysis — stories about files analysis in this arenaFiles analysis

                Stories about files analysis in this arena

                Analysis

                1. power-userHave the assistant write and run code on my data to produce charts, computed answers, and downloadable files

                  weight 3 · round drawn
                  ChatGPTfullcommunity8/10

                  ChatGPT's Code Interpreter/Data Analysis capability (documented via file drafting, spreadsheet/PDF handling, interactive visualizations, and downloadable finished files) lets it write and execute code against uploaded data to produce charts and computed answers, then export results as documents/spreadsheets/PDFs. Docs explicitly describe analyzing files, building interactive visualizations, and downloading completed files, and community evidence corroborates real code/file execution capability (e.g., running code/containers, file interpretation plugins). Missing for 10: no dedicated first-party doc page specifically titled 'Code Interpreter/Data Analysis' in this pack, and no hands-on example showing a specific chart-from-CSV walkthrough.

                  • [claimed-docs] “Draft and refine documents, presentations, spreadsheets, and PDF files. Review the result, ask for specific changes, and download the comple…”
                  • [claimed-docs] “make interactive [visualizations], and build or share websites and apps with [Sites]”
                  • [claimed-docs] “Create a deck, analyze files, draft a report, build a project plan”
                  • [claimed-docs] “Search the web, browse websites, compare sources, read files, analyze data, and summarize findings.”
                  • [claimed-docs] “Draft and refine [documents, presentations, spreadsheets, and PDF files]. Review the result, ask for specific changes, and download the comp…”
                  • [community] “This is huge, essentially adding what people have been building with LangChain Tools into the core product. The browser and file-upload/inte…”
                  proof ↗⚑ flag
                  Claudefullclaimed8/10

                  Docs explicitly describe generating downloadable Excel/PowerPoint/Word/PDF files, producing reports with charts and visualizations, and using artifacts to build code-driven visualizations and interactive components from uploaded data files (CSV, XLSX, etc.). This directly matches the power-user story of writing/running code on data to produce charts, computed answers, and downloadable outputs. missing for 10: independent/hands-on community corroboration specifically validating the data-analysis/code-execution-to-chart workflow (community evidence in the pack focuses on coding agent quality, not this analysis feature).

                  • [claimed-docs] “Prompt Claude using natural language to generate Excel spreadsheets, PowerPoint presentations, Word documents, and PDF files that you can do…”
                  • [claimed-docs] “produce reports with charts and visualizations, and generate presentations from your documents—all without specialized software skills”
                  • [claimed-docs] “Generate code and visualize data”
                  • [claimed-docs] “Claude can work with the following document types: - PDF - DOCX - CSV - TXT - HTML - ODT - RTF - EPUB - JSON - XLSX”
                  • [claimed-docs] “Claude can work with the following document types: PDF DOCX CSV TXT HTML ODT RTF EPUB JSON XLSX”
                  • [claimed-docs] “Common examples of artifact content include: Documents (Markdown or plain text) - Code snippets ... Interactive React components”
                  • [claimed-docs] “Common examples of artifact content include: - Documents (Markdown or plain text) - Code snippets - Single-page HTML websites - SVG images -…”
                  • [claimed-docs] “Artifacts allow you to turn ideas into shareable apps, tools, or content—build tools, visualizations, and experiences by simply describing w…”
                  • [claimed-docs] “build tools, visualizations, and experiences by simply describing what you need”
                  • [claimed-docs] “Files can be uploaded to individual chats or uploaded to a project's Files section for persistent reference across conversations.”
                  proof ↗⚑ flag

                Artifacts

                1. knowledge-workerHave the assistant create and iteratively edit documents, presentations, and other files I can export

                  weight 2 · round drawn
                  ChatGPTfullclaimed9/10

                  Docs strongly support creating and iteratively editing documents, presentations, spreadsheets, and PDFs, with review/annotation and export/download workflows (chatgpt-docs-8, chatgpt-docs-30, chatgpt-docs-60, chatgpt-docs-66, chatgpt-docs-68, chatgpt-docs-70), plus a dedicated artifacts/canvas viewer supporting annotations and comments for iterative refinement (chatgpt-docs-7, chatgpt-docs-18, chatgpt-docs-24, chatgpt-docs-101). Missing for 10: independent hands-on corroboration specifically of document/presentation export workflows (community evidence pack focuses on other capabilities like coding/search, not file export).

                  • [claimed-docs] “Draft and refine documents, presentations, spreadsheets, and PDF files.”
                  • [claimed-docs] “Draft and refine documents, presentations, spreadsheets, and PDF files. Review the result, ask for specific changes, and download the comple…”
                  • [claimed-docs] “Draft and refine [documents, presentations, spreadsheets, and PDF files]. Review the result, ask for specific changes, and download the comp…”
                  • [claimed-docs] “Create a deck, analyze files, draft a report, build a project plan”
                  • [claimed-docs] “Turn research and analysis into documents, presentations, spreadsheets, and other finished work.”
                  • [claimed-docs] “Use ChatGPT Work to turn notes, docs, research, or meeting materials into a structured deck.”
                  • [claimed-docs] “Annotations let you point to a specific part of a file and tell ChatGPT what to change.”
                  • [claimed-docs] “a document editor can provide tools to find a section or add a comment”
                  • [claimed-docs] “Use annotations to point at a specific part of a supported preview and request a focused revision.”
                  proof ↗⚑ flag
                  Claudefullclaimed9/10

                  Docs explicitly describe generating and editing Excel, PowerPoint, Word, and PDF files via natural-language prompts, plus Artifacts for documents/code/interactive content that can be iteratively refined and downloaded, and uploading/persisting files in Projects for ongoing editing. Missing for 10: independent hands-on review confirming export fidelity/iteration quality across file types.

                  • [claimed-docs] “Prompt Claude using natural language to generate Excel spreadsheets, PowerPoint presentations, Word documents, and PDF files that you can do…”
                  • [claimed-docs] “produce reports with charts and visualizations, and generate presentations from your documents—all without specialized software skills”
                  • [claimed-docs] “Common examples of artifact content include: Documents (Markdown or plain text) - Code snippets ... Interactive React components”
                  • [claimed-docs] “Common examples of artifact content include: - Documents (Markdown or plain text) - Code snippets - Single-page HTML websites - SVG images -…”
                  • [claimed-docs] “Files can be uploaded to individual chats or uploaded to a project's Files section for persistent reference across conversations.”
                  • [claimed-docs] “Claude can work with the following document types: - PDF - DOCX - CSV - TXT - HTML - ODT - RTF - EPUB - JSON - XLSX”
                  proof ↗⚑ flag

                Files

                1. knowledge-workerUpload documents, spreadsheets, and PDFs and get accurate analysis of their contents

                  weight 3 · round to Claude
                  ChatGPTfullclaimed7/10

                  Docs confirm document/spreadsheet/PDF upload and analysis (draft/refine documents, presentations, spreadsheets, PDFs; 'analyze files'; 'read files, analyze data, and summarize findings'), which directly matches the story. However, evidence is entirely first-party documentation with no independent hands-on validation of accuracy specifically for uploaded file analysis (community evidence covers unrelated web-search accuracy issues, not file analysis). Missing for 10: independent/hands-on verification of analysis accuracy on real uploaded documents/spreadsheets/PDFs, and detail on limits (file size, complex spreadsheet formulas, OCR quality).

                  • [claimed-docs] “Draft and refine documents, presentations, spreadsheets, and PDF files.”
                  • [claimed-docs] “Draft and refine documents, presentations, spreadsheets, and PDF files. Review the result, ask for specific changes, and download the comple…”
                  • [claimed-docs] “Draft and refine [documents, presentations, spreadsheets, and PDF files]. Review the result, ask for specific changes, and download the comp…”
                  • [claimed-docs] “Create a deck, analyze files, draft a report, build a project plan”
                  • [claimed-docs] “Search the web, browse websites, compare sources, read files, analyze data, and summarize findings.”
                  proof ↗⚑ flag
                  Claudefullclaimed9/10

                  Docs confirm Claude supports uploading PDFs, DOCX, CSV, XLSX, TXT, HTML, ODT, RTF, EPUB, JSON, with analysis of both text and visual elements in PDFs up to 100 pages, plus persistent Project file storage for cross-conversation reference and generation of derived documents/reports. Missing for 10: independent hands-on benchmarking of analysis accuracy across large/complex spreadsheets or long PDFs beyond the 100-page limit.

                  • [claimed-docs] “Claude can work with the following document types: - PDF - DOCX - CSV - TXT - HTML - ODT - RTF - EPUB - JSON - XLSX”
                  • [claimed-docs] “Claude can work with the following document types: PDF DOCX CSV TXT HTML ODT RTF EPUB JSON XLSX”
                  • [claimed-docs] “Claude analyzes both text and visual elements (like images, charts, and graphics) in PDFs of 100 pages or fewer”
                  • [claimed-docs] “Files can be uploaded to individual chats or uploaded to a project's Files section for persistent reference across conversations.”
                  • [claimed-docs] “Prompt Claude using natural language to generate Excel spreadsheets, PowerPoint presentations, Word documents, and PDF files that you can do…”
                  • [claimed-docs] “produce reports with charts and visualizations, and generate presentations from your documents—all without specialized software skills”
                  proof ↗⚑ flag

                Memory context — stories about memory context in this arenaMemory context

                Stories about memory context in this arena

                Memory

                1. power-userHave the assistant remember relevant context from previous chats and apply it in new conversations

                  weight 3 · round drawn
                  ChatGPTfullclaimed8/10

                  ChatGPT's Memories feature explicitly carries useful context from earlier chats into future conversations, with dedicated docs on memory controls (personalize.md, customization/memories.md) and even CLI-level controls (/memories) for whether a chat can use or contribute to memories. Missing for 10: independent/hands-on community corroboration of cross-chat memory recall working reliably, and detail on limits/scope of what memories retain.

                  • [claimed-docs] “Memories let ChatGPT and Codex carry useful context from earlier work into future work.”
                  • [claimed-docs] “Memories let ChatGPT carry useful context from earlier chats into future work.”
                  • [claimed-docs] “In Codex CLI, use /memories in an interactive session to control whether the current chat can use existing local memories or become an input…”
                  • [claimed-docs] “use `/memories` to choose whether a chat can use local memories or contribute to future memories”
                  proof ↗⚑ flag
                  Claudefullclaimed8/10

                  Claude has explicit first-party memory/chat-search docs: it can search previous conversations and "remember context from your chats and carry it into new conversations and Cowork tasks," plus a dedicated article on how memory works, what's remembered, and how to review/edit it; the product page also advertises "Memory across conversations" as a core feature. This directly matches the story of remembering context across chats and applying it in new ones. Missing for 10: independent/hands-on corroboration of memory quality or limitations in practice beyond vendor docs.

                  • [claimed-docs] “You can prompt Claude to search through your previous conversations to find and reference relevant information in new chats.”
                  • [claimed-docs] “Chat on web, iOS, Android, and on your desktop ... Memory across conversations”
                  • [claimed-docs] “This article explains how chat search and memory work, what Claude does and doesn’t remember, how to review and edit what’s saved, and how t…”
                  • [claimed-docs] “Claude can also remember context from your chats and carry it into new conversations and Cowork tasks.”
                  proof ↗⚑ flag

                2. power-userSet persistent custom instructions and preferences that shape every response

                  weight 1 · round to ChatGPT
                  ChatGPTfullclaimed8/10

                  ChatGPT supports persistent custom instructions/personalization and Memories that carry context across chats, plus Codex's global AGENTS.md for persistent personal instructions in coding workflows, and fine-grained control via /memories. Missing for 10: independent hands-on verification that custom instructions reliably shape *every* response over long-term use, and no detail on limits/scope conflicts between memories and per-chat overrides.

                  • [claimed-docs] “Memories let ChatGPT and Codex carry useful context from earlier work into future work.”
                  • [claimed-docs] “Memories let ChatGPT carry useful context from earlier chats into future work.”
                  • [claimed-docs] “In Codex, these personal instructions are stored in your global `AGENTS.md` file.”
                  • [claimed-docs] “use `/memories` to choose whether a chat can use local memories or contribute to future memories”
                  • [claimed-docs] “In Codex CLI, use /memories in an interactive session to control whether the current chat can use existing local memories or become an input…”
                  proof ↗⚑ flag
                  Claudepartialclaimed6/10

                  Claude supports persistent memory/context (chat search and memory, Projects with knowledge bases, custom instructions implied via Skills and Projects) that carries into new conversations, but there is no explicit documented feature for setting global 'custom instructions' that shape every response the way ChatGPT's system prompt does. missing for 10: dedicated persistent custom-instructions/preferences UI applying to all chats, independent/hands-on verification that memory reliably shapes every response, and clarity on scope/limits of what's remembered.

                  • [claimed-docs] “You can prompt Claude to search through your previous conversations to find and reference relevant information in new chats.”
                  • [claimed-docs] “This article explains how chat search and memory work, what Claude does and doesn’t remember, how to review and edit what’s saved, and how t…”
                  • [claimed-docs] “Claude can also remember context from your chats and carry it into new conversations and Cowork tasks.”
                  • [claimed-docs] “Projects allow you to create self-contained workspaces with their own chat histories and knowledge bases.”
                  • [claimed-docs] “Skills teach Claude how to complete specific tasks in a repeatable way, whether that's creating documents with your company's brand guidelin…”
                  proof ↗⚑ flag

                Projects

                1. knowledge-workerOrganize related chats and files into a project or space that shares context and instructions

                  weight 2 · round drawn
                  ChatGPTfullclaimed8/10

                  ChatGPT's Projects feature explicitly lets users organize related chats, share context/instructions, and even attach local folders/files for shared context, directly matching the story. Memories also reinforce carrying context across chats. Missing for 10: independent/hands-on corroboration of the Projects feature working as described, and detail on instruction-sharing UI beyond docs claims.

                  • [claimed-docs] “Use a project to organize related chats and give ChatGPT the context it needs.”
                  • [claimed-docs] “A local project gives chats access to one or more folders on your computer, such as a collection of source files or a codebase.”
                  • [claimed-docs] “Memories let ChatGPT and Codex carry useful context from earlier work into future work.”
                  • [claimed-docs] “Memories let ChatGPT carry useful context from earlier chats into future work.”
                  proof ↗⚑ flag
                  Claudefullclaimed8/10

                  Projects docs confirm self-contained workspaces with shared chat histories and knowledge bases, persistent file uploads scoped to a project for cross-conversation reference, and memory/chat search to build on prior context. missing for 10: no evidence of custom instructions/system prompt configuration per project beyond files, and no independent/hands-on corroboration of the project workflow in practice.

                  • [claimed-docs] “Projects allow you to create self-contained workspaces with their own chat histories and knowledge bases.”
                  • [claimed-docs] “Files can be uploaded to individual chats or uploaded to a project's Files section for persistent reference across conversations.”
                  • [claimed-docs] “This article explains how chat search and memory work, what Claude does and doesn’t remember, how to review and edit what’s saved, and how t…”
                  • [claimed-docs] “Chat on web, iOS, Android, and on your desktop ... Memory across conversations”
                  proof ↗⚑ flag

                Multimodal — stories about multimodal in this arenaMultimodal

                Stories about multimodal in this arena

                Images

                1. knowledge-workerGenerate and edit images from natural-language prompts

                  weight 2 · round to ChatGPT
                  ChatGPTfullclaimed9/10

                  ChatGPT's docs explicitly cover image generation and editing from natural-language prompts, including editing via Comment/annotations, reference images, and Canvas view for reviewing multiple images, plus mobile app support for generating/transforming images and creating stickers. This directly matches the knowledge-worker story of generating and editing images conversationally. Missing for 10: no independent/hands-on community corroboration specifically for image generation quality or reliability.

                  • [claimed-docs] “Ask ChatGPT to generate or edit images.”
                  • [claimed-docs] “use Comment to add precise feedback to one or more images”
                  • [claimed-docs] “Generate or edit images, make interactive visualizations, and build or share websites and apps with Sites.”
                  • [claimed-docs] “Ask ChatGPT to generate or edit images. Use image generation for UI assets, banners, backgrounds, illustrations”
                  • [claimed-docs] “Switch between Focused view to inspect one image and Canvas view to see the images generated in the same chat.”
                  • [claimed-docs] “In Canvas view, use Comment to add precise feedback to one or more images.”
                  • [claimed-docs] “Add a reference image when you want ChatGPT to transform an existing asset or use it as visual guidance.”
                  • [claimed-docs] “Image generation–Generate original images from a description, or transform existing ones with a few simple words.”
                  • [claimed-docs] “Choose an image and style, remix with emojis, and make personalized stickers that are ready to share in your group chat.”
                  proof ↗⚑ flag
                  Claudepartialclaimed4/10

                  Claude can understand and analyze uploaded/pasted images (JPEG, PNG, GIF, WebP) and can produce SVG images, diagrams, and flowcharts as artifacts from natural-language descriptions, but there is no evidence of true raster image generation or photo-editing capability comparable to dedicated image models. Editing of generated visual artifacts is possible via iterative prompting, but this is limited to code-rendered graphics rather than general image generation/editing. Missing for 10: dedicated raster image-generation model, photo editing/inpainting features, and any independent confirmation of image-generation quality.

                  • [claimed-docs] “Common examples of artifact content include: - Documents (Markdown or plain text) - Code snippets - Single-page HTML websites - SVG images -…”
                  • [claimed-docs] “Claude analyzes both text and visual elements (like images, charts, and graphics) in PDFs of 100 pages or fewer”
                  • [claimed-docs] “Claude supports the following image formats: - JPEG - PNG - GIF - WebP”
                  • [claimed-docs] “You can also copy images and paste them from your clipboard into Claude”
                  • [claimed-docs] “build tools, visualizations, and experiences by simply describing what you need”
                  proof ↗⚑ flag

                2. knowledge-workerShare screenshots and photos and have the assistant accurately interpret what is in them

                  weight 2 · round drawn
                  ChatGPTfullclaimed7/10

                  First-party docs confirm ChatGPT accepts photo/screenshot uploads and interprets their content (e.g., transcribing a handwritten recipe or identifying a landmark from a photo), which directly matches the knowledge-worker use case of sharing images for analysis. Missing for 10: independent/hands-on evidence corroborating accuracy of image interpretation, and explicit documentation of screenshot-specific analysis (e.g., UI screenshots) rather than just general photo uploads.

                  • [claimed-docs] “Photo upload—Snap or upload a picture to transcribe a handwritten recipe or get info about a landmark.”
                  • [claimed-docs] “Snap or upload a picture to transcribe a handwritten recipe or get info about a landmark.”
                  • [claimed-docs] “Image generation–Generate original images from a description, or transform existing ones with a few simple words.”
                  proof ↗⚑ flag
                  Claudefullclaimed7/10

                  Claude's docs confirm native image upload/paste support (JPEG, PNG, GIF, WebP) and clipboard paste, plus PDF analysis that includes visual elements like images and charts, directly supporting screenshot/photo interpretation for knowledge workers. Missing for 10: independent/hands-on evidence validating accuracy of image interpretation, and no explicit mention of screenshot-specific use cases (e.g., UI screenshots, photos of documents) beyond general image/PDF support.

                  • [claimed-docs] “Claude analyzes both text and visual elements (like images, charts, and graphics) in PDFs of 100 pages or fewer”
                  • [claimed-docs] “Claude supports the following image formats: - JPEG - PNG - GIF - WebP”
                  • [claimed-docs] “You can also copy images and paste them from your clipboard into Claude”
                  • [claimed-docs] “Claude can work with the following document types: - PDF - DOCX - CSV - TXT - HTML - ODT - RTF - EPUB - JSON - XLSX”
                  proof ↗⚑ flag

                Voice

                1. knowledge-workerHave a natural, real-time voice conversation with the assistant

                  weight 2 · round drawn
                  ChatGPTfullclaimed8/10

                  ChatGPT ships Advanced Voice Mode for real-time spoken conversation on mobile ('Tap the soundwave icon to have a real-time convo on the go') and ChatGPT Voice on desktop that lets users talk through ideas, start work, check progress, or change direction without switching to typing. This directly matches the knowledge-worker story of natural, real-time voice conversation. Missing for 10: independent hands-on corroboration of voice quality/latency and any community verification beyond vendor docs.

                  • [claimed-docs] “Advanced Voice Mode–Tap the soundwave icon to have a real-time convo on the go.”
                  • [claimed-docs] “Advanced Voice Mode–Tap the soundwave icon to have a real-time convo on the go. Settle a dinner table debate, or practice a new language.”
                  • [claimed-docs] “ChatGPT Voice lets you talk through ideas and coordinate tasks in Chat, Work, and Codex in the ChatGPT desktop app. Start work, check progre…”
                  • [claimed-docs] “Start work, check progress, or change direction without switching back to typing.”
                  • [claimed-docs] “ChatGPT Voice lets you talk through ideas and coordinate tasks in Chat, Work, and Codex in the ChatGPT desktop app.”
                  • [claimed-docs] “ChatGPT Voice can start separate tasks for longer work, check existing tasks, and send follow-up instructions.”
                  proof ↗⚑ flag
                  Claudefullclaimed8/10

                  Claude ships an explicit Voice mode enabling complete spoken conversations, available across web, desktop, iOS and Android, positioned to work best on phone. Missing for 10: independent hands-on reviews of voice latency/naturalness and confirmation it's out of beta.

                  • [claimed-docs] “Voice mode allows you to have complete spoken conversations with Claude.”
                  • [claimed-docs] “Voice mode is a beta feature available to all plans (Free, Pro, Max, Team, and Enterprise) on Claude Mobile (iOS and Android), Claude Deskto…”
                  proof ↗⚑ flag

                Openness — open source, data portability, and self-hosting storiesOpenness

                Open source, data portability, and self-hosting stories

                1. ai-native userDo everything through the API that I can do in the UI

                  weight 2 · round to ChatGPT
                  ChatGPTpartialprobed3/10

                  There is evidence of programmatic access via the Codex CLI/SDK for coding automation and CI jobs, and a security-scan CLI, but nothing shows that the broad set of UI-only capabilities (Canvas/image editing, Computer Use, Voice, Work desktop automations, Sites, MCP-in-chat, memories, scheduled tasks) are exposed through a general ChatGPT API with parity to the UI. A probe for an OpenAPI/swagger spec on the docs site also returned 404s, reinforcing the lack of a documented full API surface. missing for 10: documented API endpoints mirroring Canvas/image tools, Computer Use, Voice, Work automations, Sites, and MCP-in-chat; independent confirmation of API-UI feature parity.

                  • [claimed-docs] “Use the Codex SDK to automate coding tasks, including jobs in CI.”
                  • [claimed-docs] “Inspect code, make changes, run commands, and automate repeatable work without leaving your terminal.”
                  • [claimed-docs] “Use its command-line interface (CLI) to scan repositories you own or have permission to assess, review findings over time”
                  • [probe] “PROBE openapi: all candidate paths 404 (https://learn.chatgpt.com/openapi.json, https://learn.chatgpt.com/swagger.json, https://learn.chatgp…”
                  proof ↗⚑ flag
                  Claudenone0/10

                  The evidence pack details many UI-exclusive Claude.ai features (Cowork, voice mode, computer use, memory/chat search, artifacts, browser extension, scheduled tasks) but contains no documentation that these capabilities are exposed through the Claude API, nor any statement of API/UI feature parity. Enterprise API mentions are limited to a Compliance API for logs, not general feature parity.

                  • [claimed-docs] “With Cowork, you can describe an outcome, step away, and come back to finished work—formatted documents, organized files, synthesized resear…”
                  • [claimed-docs] “Voice mode allows you to have complete spoken conversations with Claude.”
                  • [claimed-docs] “Claude opens sites, reads pages, clicks, types, and fills forms while you watch, with no need to switch windows.”
                  • [claimed-docs] “it may navigate to your screen directly—clicking, typing, and opening apps just like you would.”
                  • [claimed-docs] “Audit logs: capture key information about user actions, system events, and data access. ... Compliance API: programmatically access Claude u…”
                  • [claimed-docs] “Projects allow you to create self-contained workspaces with their own chat histories and knowledge bases.”
                  • [claimed-docs] “This article explains how chat search and memory work, what Claude does and doesn’t remember, how to review and edit what’s saved, and how t…”
                  proof ↗⚑ flag

                2. ai-native userExport all of my data in open formats and leave

                  weight 3 · round to Claude
                  ChatGPTnone0/10

                  No evidence in the pack of a data export feature producing open/portable formats, nor any mention of account data export or deletion workflow. Evidence only covers general product features (Codex, Work, MCP, image gen, etc.) with nothing about exporting personal chat/data history or leaving the platform with your data intact.

                    ⚑ flag
                    Claudepartialclaimed6/10

                    Claude documents a data-export feature covering conversation and account data, letting users take their data with them ([claude-docs-12], [claude-docs-40]). However, there is no evidence specifying the export format is an open/standard one (e.g., JSON/portable), nor documentation of easy migration/interoperability with other tools, so the 'open format' and full portability aspects of the story are unconfirmed. missing for 10: explicit statement of open/standard export format, evidence of full interoperability/reuse elsewhere, and independent confirmation of export completeness.

                    • [claimed-docs] “Individual Claude users can export user information and chat history from Settings > Privacy”
                    • [claimed-docs] “Data exports include conversation data and the user data for your account.”
                    proof ↗⚑ flag

                  Privacy posture — data-handling and privacy storiesPrivacy posture

                  Data-handling and privacy stories

                  1. ai-native userChoose where my data is stored (region/residency)

                    weight 2 · round drawn
                    ChatGPTnone0/10

                    No evidence in the pack addresses data residency or regional storage choice for ChatGPT; the evidence only touches general data controls (e.g., Computer Use screenshots) without any region-selection capability.

                      ⚑ flag
                      Claudenone0/10

                      No evidence in the pack mentions data residency, regional storage options, or geographic controls for where account/chat data is stored; the closest items concern export, audit logs, and training opt-outs, none of which address residency choice.

                        ⚑ flag

                      • ai-native userPrevent my data from being used to train AI models

                        weight 3 · round to Claude
                        ChatGPTnone0/10

                        The evidence pack contains no documentation or mention of data-training opt-out controls, privacy settings, or 'improve the model for everyone' toggles; only tangential data-controls mention (chatgpt-docs-89) refers to Computer Use screenshots, not training opt-out generally. Missing for 10: any reference to training-data opt-out settings, business/API data-usage policies, or privacy dashboard controls.

                        • [claimed-docs] “Your ChatGPT data controls apply to content processed through ChatGPT, including screenshots taken by Computer Use.”
                        proof ↗⚑ flag
                        Claudefullclaimed8/10

                        Claude documents explicit user controls to prevent training use: incognito chats are never used to improve Claude even with Model Improvement enabled, users can toggle the 'Model Improvement' privacy setting, and connector data (Gmail/Drive/Calendar) is explicitly excluded from training. This directly satisfies the ai-native privacy-posture story of preventing data from being used for model training. Missing for 10: independent/third-party verification of these claims and explicit default policy for Enterprise/Team plans beyond connectors.

                        • [claimed-docs] “Your incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your privacy settings.”
                        • [claimed-docs] “When you allow us to use your chats or coding sessions to help improve Claude, we implement several layers of protection for your privacy.”
                        • [claimed-docs] “We do not train our models on your Gmail, Drive, or Calendar connector data, ensuring your private information remains private.”
                        proof ↗⚑ flag

                      • ai-native userControl data retention and deletion

                        weight 2 · round to Claude
                        ChatGPTpartialclaimed3/10

                        The only relevant evidence is a passing reference that 'ChatGPT data controls apply to content processed through ChatGPT, including screenshots taken by Computer Use,' implying some data-control/retention settings exist, but no documentation details how to view, export, or delete data, set retention periods, or manage memory deletion. Missing for 10: explicit data retention/export/delete documentation, memory deletion controls, and independent confirmation that these controls work as described.

                        • [claimed-docs] “Your ChatGPT data controls apply to content processed through ChatGPT, including screenshots taken by Computer Use.”
                        proof ↗⚑ flag
                        Claudefullclaimed8/10

                        Claude provides explicit user-facing data controls: exporting chat history/user data (claude-docs-12, 40), reviewing/editing/turning off memory (claude-docs-45), incognito chats excluded from training (claude-docs-46), and enterprise audit logs plus a compliance API for data governance (claude-docs-11, 44). Missing for 10: explicit self-service account/data deletion flow documentation and independent verification that deletion requests are honored.

                        • [claimed-docs] “Individual Claude users can export user information and chat history from Settings > Privacy”
                        • [claimed-docs] “Data exports include conversation data and the user data for your account.”
                        • [claimed-docs] “This article explains how chat search and memory work, what Claude does and doesn’t remember, how to review and edit what’s saved, and how t…”
                        • [claimed-docs] “Your incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your privacy settings.”
                        • [claimed-docs] “When you allow us to use your chats or coding sessions to help improve Claude, we implement several layers of protection for your privacy.”
                        • [claimed-docs] “Audit logs: capture key information about user actions, system events, and data access. ... Compliance API: programmatically access Claude u…”
                        • [claimed-docs] “Enterprise includes everything in the Team plan, plus the following: - Security features to ensure the safety and compliance of your organiz…”
                        proof ↗⚑ flag

                      • ai-native userOpt out of telemetry and usage tracking

                        weight 2 · round to Claude
                        ChatGPTnone0/10

                        None of the evidence addresses telemetry/usage-tracking opt-out controls for ChatGPT; the docs cover data controls tangentially (e.g., data usage for Computer Use screenshots) but nothing about disabling telemetry or usage analytics.

                          ⚑ flag
                          Claudepartialclaimed6/10

                          Claude documents a 'Model Improvement' privacy setting that can be toggled off and 'incognito chats' that are excluded from training even if Model Improvement is enabled, giving users control over whether their conversations are used to improve the model (claude-docs-46, claude-docs-47). This addresses opt-out of usage-for-training, and data export/deletion options exist (claude-docs-12, claude-docs-40), but there is no explicit documentation of a broader telemetry/analytics opt-out (e.g., product usage metrics, crash reporting) beyond model-training data use. missing for 10: explicit telemetry/analytics tracking opt-out settings, independent confirmation that toggling actually stops all usage tracking, documentation of what non-training telemetry data is collected.

                          • [claimed-docs] “Your incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your privacy settings.”
                          • [claimed-docs] “When you allow us to use your chats or coding sessions to help improve Claude, we implement several layers of protection for your privacy.”
                          • [claimed-docs] “Individual Claude users can export user information and chat history from Settings > Privacy”
                          • [claimed-docs] “Data exports include conversation data and the user data for your account.”
                          proof ↗⚑ flag

                        Research answers — stories about research answers in this arenaResearch answers

                        Stories about research answers in this arena

                        Research

                        1. knowledge-workerLaunch a deep research run that autonomously searches many sources and returns a cited report

                          weight 3 · round to Claude
                          ChatGPTpartialclaimed6/10

                          Docs show ChatGPT can search the web, browse multiple sources, compare them, and produce cited results in-chat (docs-86, docs-90), and can turn research into finished documents/reports (docs-66, docs-68, docs-60). However, there is no explicit mention of a dedicated 'Deep Research' autonomous multi-source research mode/run with a structured long-form cited report as its own distinct feature — the evidence only shows general web-search-with-citations and document drafting capabilities. missing for 10: an explicitly named deep-research mode, evidence of autonomous long-running multi-source research runs, and a structured final cited-report output format.

                          • [claimed-docs] “Search results and citations appear in the chat when ChatGPT uses web search.”
                          • [claimed-docs] “Search the web, browse websites, compare sources, read files, analyze data, and summarize findings.”
                          • [claimed-docs] “Create a deck, analyze files, draft a report, build a project plan”
                          • [claimed-docs] “Turn research and analysis into documents, presentations, spreadsheets, and other finished work.”
                          • [claimed-docs] “Draft and refine [documents, presentations, spreadsheets, and PDF files]. Review the result, ask for specific changes, and download the comp…”
                          • [claimed-docs] “In the CLI, pass `--search` to fetch live results for one run”
                          proof ↗⚑ flag
                          Claudefullcommunity8/10

                          Claude's Research feature is well documented: it operates agentically running multiple searches that build on each other, determines what to investigate next, and delivers thorough answers in minutes with easy-to-check citations. Community evidence is mixed on research depth (some say it lags ChatGPT/Gemini), which caps quality but the core capability is clearly delivered. Missing for 10: independent hands-on benchmark showing citation accuracy/report quality, and clearer detail on breadth of sources searched.

                          • [claimed-docs] “With research, Claude delivers thorough answers in minutes, complete with easy-to-check citations so you can trust Claude's findings.”
                          • [claimed-docs] “Research transforms how Claude finds and analyzes information. Claude operates agentically, conducting multiple searches that build on each …”
                          • [claimed-docs] “Claude operates agentically, conducting multiple searches that build on each other while determining exactly what to investigate next.”
                          • [claimed-docs] “Claude delivers thorough answers in minutes, complete with easy-to-check citations so you can trust Claude's findings.”
                          • [community] “Meanwhile, Claude's general use cases are... fine. For generic research topics, I find that ChatGPT and Gemini run circles around it: in the…”
                          • [community] “Works pretty nicely for research still, not seeing a substantial qualitative improvement over Opus 4.5.”
                          proof ↗⚑ flag

                        2. knowledge-workerGet answers grounded in current web results with citations back to the sources

                          weight 2 · round to ChatGPT
                          ChatGPTfullcommunity8/10

                          ChatGPT's web search docs explicitly state that search results and citations appear in chat when web search is used, and describe searching, browsing, comparing sources, and summarizing findings — directly matching the story. Community feedback shows mixed satisfaction with search accuracy (e.g., comparisons to Perplexity, a weather inaccuracy) but doesn't concretely show citations failing to appear, so this doesn't rise to a dispute. Missing for 10: independent verification of citation accuracy/consistency across many queries and no first-party detail on citation formatting/source diversity.

                          • [claimed-docs] “Search results and citations appear in the chat when ChatGPT uses web search.”
                          • [claimed-docs] “Search the web, browse websites, compare sources, read files, analyze data, and summarize findings.”
                          • [claimed-docs] “In the CLI, pass `--search` to fetch live results for one run”
                          • [community] “AKA Bing Search in ChatGPT. So it is not it's own search engine and is still using Bing for its results just like the rest of them.”
                          • [community] “I gave it a quick spin and my initial impression is much worse than perplexity.”
                          • [community] “I asked it the current weather in my area and the temperature was off by 23 degrees F.”
                          proof ↗⚑ flag
                          Claudepartialcommunity7/10

                          Claude's Research feature is documented to perform agentic multi-step web searches and deliver answers with 'easy-to-check citations' (claude-docs-20, 27, 39, 68), directly matching the story. However, independent community feedback suggests mixed real-world quality—commenters say Claude's general research 'is fine' but that ChatGPT and Gemini 'run circles around it' in depth and presentation (claude-comm-8, claude-comm-9), tempering confidence in how well-grounded/comprehensive the citations truly are. missing for 10: independent hands-on verification of citation accuracy/source quality, and confirmation research draws from live/current web data versus stale index.

                          • [claimed-docs] “With research, Claude delivers thorough answers in minutes, complete with easy-to-check citations so you can trust Claude's findings.”
                          • [claimed-docs] “Research transforms how Claude finds and analyzes information. Claude operates agentically, conducting multiple searches that build on each …”
                          • [claimed-docs] “Claude operates agentically, conducting multiple searches that build on each other while determining exactly what to investigate next.”
                          • [claimed-docs] “Claude delivers thorough answers in minutes, complete with easy-to-check citations so you can trust Claude's findings.”
                          • [community] “Meanwhile, Claude's general use cases are... fine. For generic research topics, I find that ChatGPT and Gemini run circles around it: in the…”
                          • [community] “Works pretty nicely for research still, not seeing a substantial qualitative improvement over Opus 4.5.”
                          proof ↗⚑ flag

                        Trust controls — stories about trust controls in this arenaTrust controls

                        Stories about trust controls in this arena

                        Admin

                        1. team-adminManage members, permissions, and data policies for my organization's workspace

                          weight 2 · round drawn
                          ChatGPTpartialclaimed4/10

                          Only a single explicit doc line notes that a workspace administrator can control access to ChatGPT Work, plugins, browser use, and network access, plus general mentions of GPT builders using 'approved workspace apps' and data controls applying to processed content. There is no evidence of a full admin console covering member invitation/removal, granular role-based permissions, or explicit data retention/training-opt-out policy controls. missing for 10: admin console/member management UI, granular role/permission settings, explicit data-retention and training-opt-out policy controls, independent corroboration of these admin features.

                          • [claimed-docs] “Your workspace administrator can control access to ChatGPT Work, plugins, browser use, and network access.”
                          • [claimed-docs] “Allow GPT builders to use approved workspace apps or configure actions that interact with permitted third-party APIs.”
                          • [claimed-docs] “configure actions that interact with permitted third-party APIs”
                          • [claimed-docs] “Your ChatGPT data controls apply to content processed through ChatGPT, including screenshots taken by Computer Use.”
                          proof ↗⚑ flag
                          Claudepartialclaimed4/10

                          Evidence confirms Enterprise plan admin/security features like audit logs and a Compliance API for programmatic access to usage data, implying some org-level governance, but there is no documentation of member management, role/permission assignment, or granular data policy controls for team admins. missing for 10: member invitation/removal workflows, role-based permission management, workspace-level data retention/policy settings, and independent corroboration of admin console functionality.

                          • [claimed-docs] “Audit logs: capture key information about user actions, system events, and data access. ... Compliance API: programmatically access Claude u…”
                          • [claimed-docs] “Audit logs: capture key information about user actions, system events, and data access.”
                          • [claimed-docs] “Enterprise includes everything in the Team plan, plus the following: - Security features to ensure the safety and compliance of your organiz…”
                          proof ↗⚑ flag

                        Data controls

                        1. knowledge-workerExport my complete chat history and account data

                          weight 1 · round to Claude
                          ChatGPTnone0/10

                          The evidence pack contains no mention of data export, account data download, or chat history export features anywhere in the docs, community, or probe results; there is only a passing reference to 'data controls' applying to Computer Use content, which does not address exporting complete chat history or account data. Missing for 10: any documentation of an export data feature, its scope (chats, files, settings), format, or process, and any independent corroboration it works.

                          • [claimed-docs] “Your ChatGPT data controls apply to content processed through ChatGPT, including screenshots taken by Computer Use.”
                          proof ↗⚑ flag
                          Claudefullclaimed9/10

                          Claude's docs explicitly state individual users can export user information and chat history from Settings > Privacy, and that exports include both conversation data and account/user data; Enterprise adds a Compliance API for programmatic access to chat histories and file content. Missing for 10: independent/hands-on confirmation of export completeness or format details beyond first-party docs.

                          • [claimed-docs] “Individual Claude users can export user information and chat history from Settings > Privacy”
                          • [claimed-docs] “Data exports include conversation data and the user data for your account.”
                          • [claimed-docs] “Audit logs: capture key information about user actions, system events, and data access. ... Compliance API: programmatically access Claude u…”
                          proof ↗⚑ flag

                        2. knowledge-workerControl whether my conversations are used to train models

                          weight 3 · round to Claude
                          ChatGPTnone0/10

                          The evidence pack covers ChatGPT's agentic/feature capabilities (Codex, Work, MCP, Computer Use, etc.) but contains no documentation or mention of data controls, training opt-out settings, or 'Improve the model for everyone' toggles that let a user control whether their conversations are used for model training.

                            ⚑ flag
                            Claudefullclaimed8/10

                            Claude provides explicit privacy controls: incognito chats are excluded from model training even with Model Improvement enabled, a 'Model Improvement' opt-in/out toggle exists, and connector data (Gmail/Drive/Calendar) is explicitly excluded from training. Missing for 10: independent/hands-on verification that the training opt-out is actually honored in practice, and clearer documentation of the toggle's exact location/scope for all plan tiers.

                            • [claimed-docs] “Your incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your privacy settings.”
                            • [claimed-docs] “When you allow us to use your chats or coding sessions to help improve Claude, we implement several layers of protection for your privacy.”
                            • [claimed-docs] “We do not train our models on your Gmail, Drive, or Calendar connector data, ensuring your private information remains private.”
                            proof ↗⚑ flag

                          Not comparable on these axes

                          1. ai-native userConnect an agent via an official MCP server

                            weight 3 · not comparable
                            ChatGPTfullclaimed8/10

                            ChatGPT/Codex has official first-party documentation for MCP support: users can add official MCP servers (e.g., Context7 for docs, Figma, browser tools) to extend the agent, configuration is stored in config.toml, servers can be scoped per-project, and setup carries across clients without redoing it. This directly evidences 'connecting an agent via an official MCP server.' Missing for 10: independent/hands-on community corroboration of MCP server usage and broader detail on the range of officially supported/verified servers beyond the Context7 example.

                            • [claimed-docs] “let it interact with developer tools like your browser or Figma”
                            • [claimed-docs] “Once you configure your MCP servers, you can switch among those clients without redoing setup.”
                            • [claimed-docs] “Use it to give ChatGPT or Codex access to third-party documentation, or to let it interact with developer tools like your browser or Figma.”
                            • [claimed-docs] “to add Context7 (a free MCP server for developer documentation)”
                            • [claimed-docs] “you can also scope MCP servers to a project with `.codex/config.toml` (trusted projects only)”
                            • [claimed-docs] “Codex stores MCP configuration in `config.toml`”
                            proof ↗⚑ flag
                            Clauden/a

                            Claude is itself an AI agent/assistant (client role); the evidence only shows Claude connecting to or building remote MCP servers as a client (claude-docs-18, claude-docs-31, claude-docs-36, claude-probe-4), which is the separate MCP-client story. There is no evidence Claude itself runs as an MCP server that other agents could connect to, so this server-role axis does not apply to this product.

                              ⚑ flag

                            • ai-native userRead the product's source under an open license

                              weight 2 · not comparable
                              ChatGPTn/a

                              ChatGPT is a closed-source, proprietary SaaS product; open-sourcing its source code is not a plausible axis for this kind of product (unlike an open-source framework or tool), so this is a category mismatch rather than a missing capability.

                                ⚑ flag
                                Claudenone0/10

                                Claude is closed-source; there is no evidence of an open-license source release for the model or app, and community evidence even criticizes it as closed/opaque compared to FOSS alternatives like Codex CLI (claude-comm-6).

                                • [community] “Codex CLI is FOSS, unlike Claude Code, so Codex is less likely to do things like that, and it's one more reason to avoid Claude Code and Cla…”
                                proof ↗⚑ flag

                              • ai-native userSelf-host the core product

                                weight 3 · not comparable
                                ChatGPTn/a

                                ChatGPT is a closed, hosted proprietary product with no evidence of a self-hostable core model or server; self-hosting is not a fair axis for this SaaS product category (it does not ship open weights or an installable core).

                                  ⚑ flag
                                  Clauden/a

                                  Claude is a closed, hosted proprietary model/service with no self-hosting option; self-hosting the core product is a category error for this type of SaaS/AI assistant offering, not an unmet applicable axis.

                                    ⚑ flag