Groq vs Together AI

Agent access

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

   weight 2 · round drawn
   Groq✓fullprobed9/10

   Direct probe evidence confirms llms.txt is live (HTTP 200) at console.groq.com/llms.txt, and Groq also serves per-page .md docs (docs/overview.md) explicitly designed for agent consumption, indicating agent-oriented docs beyond just llms.txt. Missing for 10: no independent/community confirmation of an agent actually using llms.txt in practice.

   • [probe] “PROBE llms.txt: HTTP 200 at https://console.groq.com/llms.txt # https://console.groq.com llms.txt - [JigsawStack 🧩](https://console.groq.c…”
   • [probe] “PROBE docs-md: HTTP 200 at https://console.groq.com/docs/overview.md --- description: Fast LLM inference, OpenAI-compatible. Simple to integ…”

   proof ↗⚑ flag

   Together AI✓fullprobed9/10

   Together AI publishes a working llms.txt (HTTP 200) with a documentation index, per-page .md docs, and even a live MCP server for docs lookup, all independently confirmed via probes plus first-party docs on agent-skills. missing for 10: no independent third-party report of an agent actually consuming llms.txt successfully in practice.

   • [probe] “PROBE llms.txt: HTTP 200 at https://docs.together.ai/llms.txt # Together AI docs > Documentation for the Together AI platform for inference…”
   • [probe] “PROBE docs-md: HTTP 200 at https://docs.together.ai/intro.md > ## Documentation Index > Fetch the complete documentation index at: https://d…”
   • [probe] “PROBE mcp-endpoint (2026-09-04): POST initialize to https://docs.together.ai/mcp answered HTTP 200 with a JSON-RPC/MCP response (event: mess…”
   • [claimed-docs] “Make your AI coding agent Together-AI-aware with ready-made skills for code generation and an MCP server for live docs lookup.”
   • [claimed-docs] “Docs MCP server: Gives your agent live access to this documentation site so it can look up current information without leaving your editor.”

   proof ↗⚑ flag

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

   weight 2 · round to Groq
   Groq✓fullprobed8/10

   Groq is a pure API/inference service accessed via OpenAI-compatible REST endpoints and SDKs (API key + base_url swap), with batch processing for large-scale asynchronous automation and confirmed live, key-gated JSON endpoints — all of which are inherently scriptable/headless with no GUI dependency, ideal for CI pipelines. Missing for 10: explicit CI/CD examples (e.g., GitHub Actions) or docs specifically framing headless/automation use cases beyond batch API references.

   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”
   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”
   • [github] “Request parameters that correspond to file uploads can be passed in many different forms”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.groq.com/openai/v1/models without a key returned HTTP 401 ({"error":{"message":"Invalid …”
   • [claimed-docs] “To enable streaming, set the parameter stream=True. The completion function will then return an iterator of completion deltas”

   proof ↗⚑ flag

   Together AI✓fullprobed7/10

   Together AI is an API/SDK-first inference platform with OpenAI-compatible REST endpoints, official server-side SDKs (TypeScript/Python), batch inference, streaming via SSE, and structured/function-calling outputs — all natively usable in headless CI/automation pipelines without any UI dependency. Live probes confirm the API and status endpoints are reachable programmatically. missing for 10: no explicit CI/CD example or GitHub Actions template in evidence, no documented CLI exit-code/scripting conventions, and no independent case study of a CI pipeline using Together AI headlessly.

   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”
   • [claimed-docs] “Call 100+ open-source models with per-token pricing and no provisioning latency.”
   • [claimed-docs] “Run asynchronous batch workloads at up to 50% lower cost.”
   • [github] “We provide support for streaming responses using Server Sent Events (SSE).”
   • [github] “This library provides convenient access to the Together REST API from server-side TypeScript or JavaScript.”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.together.xyz/v1/models without a key returned HTTP 401 (Missing API key) — the OpenAI-st…”
   • [claimed-docs] “Call chat, image, audio, embedding, and more through one API.”

   proof ↗⚑ flag

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

   weight 3 · round to Groq
   Groq✓fullprobed8/10

   Groq's Responses/Chat API supports remote MCP: you provide an MCP server URL and Groq handles tool orchestration without implementing tool logic yourself, documented explicitly in first-party docs and confirmed via a probe of the dedicated MCP docs page. Missing for 10: independent hands-on user reports specifically validating MCP tool use (community evidence covers speed/reliability but not MCP), and no details on multi-server or auth edge cases.

   • [claimed-docs] “point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”
   • [claimed-docs] “you simply point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”
   • [claimed-docs] “Groq's Responses API supports remote tool use via MCP servers via HTTPS where Groq handles all orchestration... You don't implement anything…”
   • [probe] “official MCP server documented at https://console.groq.com/docs/mcp”

   proof ↗⚑ flag

   Together AI—none0/10

   The evidence only shows Together AI publishing its own MCP server (docs lookup) for coding agents to consume — i.e., Together acting as an MCP *server*, not as a client that lets users plug external MCP servers into it. There's no documentation or probe showing Together's platform/API can connect to and use tools from user-supplied MCP servers, despite having function/tool-calling APIs. missing for 10: any evidence of MCP-client support (configuring/connecting external MCP servers within Together's product), integration of MCP tool results into its agentic function-calling flow.

   • [claimed-docs] “Make your AI coding agent Together-AI-aware with ready-made skills for code generation and an MCP server for live docs lookup.”
   • [claimed-docs] “12 domain-specific skills that load on demand and teach your agent how to write correct Together AI code (right model IDs, SDK patterns, bes…”
   • [claimed-docs] “Gives your agent live access to this documentation site so it can look up current information without leaving your editor.”
   • [claimed-docs] “Docs MCP server: Gives your agent live access to this documentation site so it can look up current information without leaving your editor.”
   • [probe] “official MCP server documented at https://docs.together.ai/docs/agent-skills”
   • [probe] “PROBE mcp-endpoint (2026-09-04): POST initialize to https://docs.together.ai/mcp answered HTTP 200 with a JSON-RPC/MCP response (event: mess…”
   • [claimed-docs] “To build agent loops, chain tool calls inside one response (multi-step), and conversations that thread tools across many turns (multi-turn).”
   • [claimed-docs] “Multi-step function calling chains sequential function calls within one conversation turn.”

   proof ↗⚑ flag

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

   weight 3 · round to Together AI
   Groq~partialprobed4/10

   There is a probe hit indicating an official MCP-server doc page exists (console.groq.com/docs/mcp), but the evidence pack contains no actual content, setup instructions, or independent confirmation of this server — the detailed MCP docs quoted (groq-docs-6/17/28) describe Groq's API acting as an MCP *client* consuming remote MCP servers, not Groq serving as an MCP server for other agents to connect to. Missing for 10: quoted documentation of the official MCP server's endpoint/connection instructions, evidence of an agent successfully connecting to Groq via that server, and independent/community confirmation it works as advertised.

   • [probe] “official MCP server documented at https://console.groq.com/docs/mcp”
   • [claimed-docs] “point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”
   • [claimed-docs] “Groq's Responses API supports remote tool use via MCP servers via HTTPS where Groq handles all orchestration... You don't implement anything…”

   proof ↗⚑ flag

   Together AI✓fullprobed8/10

   Together AI documents an official Docs MCP server that gives agents live access to documentation, and a live probe confirms the MCP endpoint at docs.together.ai/mcp responds correctly to an initialize call with proper protocol capabilities. This is corroborated by both first-party docs and independent hands-on verification. Missing for 10: broader coverage beyond docs lookup (e.g., an MCP server exposing inference/model-management actions rather than just documentation retrieval), and no community discussion confirming real-world agent integration experiences.

   • [claimed-docs] “Make your AI coding agent Together-AI-aware with ready-made skills for code generation and an MCP server for live docs lookup.”
   • [claimed-docs] “12 domain-specific skills that load on demand and teach your agent how to write correct Together AI code (right model IDs, SDK patterns, bes…”
   • [claimed-docs] “Gives your agent live access to this documentation site so it can look up current information without leaving your editor.”
   • [claimed-docs] “Docs MCP server: Gives your agent live access to this documentation site so it can look up current information without leaving your editor.”
   • [probe] “official MCP server documented at https://docs.together.ai/docs/agent-skills”
   • [probe] “PROBE mcp-endpoint (2026-09-04): POST initialize to https://docs.together.ai/mcp answered HTTP 200 with a JSON-RPC/MCP response (event: mess…”

   proof ↗⚑ flag

5. ai-native userUse an official CLI

   weight 2 · round to Together AI
   Groq—none0/10

   No evidence of an official Groq CLI tool; documentation covers SDKs (Python/TypeScript), REST API, and MCP integration but no CLI is mentioned anywhere in the evidence pack.

   ⚑ flag

   Together AI~partialclaimed5/10

   Docs mention an official CLI (the 'tg' command, e.g. 'tg beta endpoints deploy...') and explicitly state fine-tuning jobs can be launched 'through the API/SDK, or with the CLI', confirming a real CLI exists for agentic/dev workflows. However, evidence lacks a dedicated CLI reference page, install instructions, full command coverage, or independent hands-on corroboration. Missing for 10: dedicated CLI documentation/installation guide, comprehensive command reference, independent user verification of CLI usage.

   • [claimed-docs] “tg beta endpoints deploy google/gemma-4-E4B-it --endpoint my-endpoint”
   • [claimed-docs] “You can launch a fine-tuning job from the console, through the API/SDK, or with the CLI”

   proof ↗⚑ flag

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

   weight 3 · round drawn
   Groq✓fullprobed9/10

   Groq ships an extensively documented, OpenAI-compatible REST API (chat completions, streaming, tool use, structured outputs, batch, service tiers, LoRA, Responses API) with a live confirmed endpoint (401 on missing key, not 404) and a public status page. Missing for 10: a discoverable formal OpenAPI/Swagger spec (all candidate paths 404).

   • [claimed-docs] “Fast LLM inference, OpenAI-compatible. Simple to integrate, easy to scale. Start building in minutes.”
   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “To start using Groq with OpenAI's client libraries, pass your Groq API key to the api_key parameter and change the base_url to https://api.g…”
   • [claimed-docs] “Generating text with Groq's Chat Completions API enables you to have natural, conversational interactions with Groq's large language models.”
   • [claimed-docs] “To use tools, the model must be provided with tool definitions. These tool definitions are in JSON schema format and are passed to the model…”
   • [claimed-docs] “Structured Outputs is a feature that ensures your model responses conform to your provided JSON Schema”
   • [claimed-docs] “Groq's Responses API is fully compatible with OpenAI's Responses API, making it easy to integrate advanced conversational AI capabilities in…”
   • [probe] “PROBE llms.txt: HTTP 200 at https://console.groq.com/llms.txt # https://console.groq.com llms.txt - [JigsawStack 🧩](https://console.groq.c…”
   • [probe] “PROBE docs-md: HTTP 200 at https://console.groq.com/docs/overview.md --- description: Fast LLM inference, OpenAI-compatible. Simple to integ…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://console.groq.com/openapi.json, https://console.groq.com/swagger.json, https://console.groq.c…”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.groq.com/openai/v1/models without a key returned HTTP 401 ({"error":{"message":"Invalid …”
   • [probe] “PROBE status-page (2026-09-04): https://groqstatus.com returns HTTP 200 and renders a public service-status page (page body includes "operat…”

   proof ↗⚑ flag

   Together AI✓fullprobed9/10

   Together AI exposes an OpenAI-compatible REST API with official SDKs (TypeScript/JS), documented endpoints for chat, batch, fine-tuning, dedicated deployment, and a live, publicly reachable models endpoint confirmed by probe; docs are extensive and public. Missing for 10: a discoverable formal OpenAPI/swagger spec (probe found all candidate paths 404).

   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”
   • [claimed-docs] “Call 100+ open-source models with per-token pricing and no provisioning latency.”
   • [github] “This library provides convenient access to the Together REST API from server-side TypeScript or JavaScript.”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.together.xyz/v1/models without a key returned HTTP 401 (Missing API key) — the OpenAI-st…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.together.ai/openapi.json, https://docs.together.ai/swagger.json, https://docs.together.…”

   proof ↗⚑ flag

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

   weight 2 · round drawn
   Groq—none0/10

   Evidence covers billing spend limits and general API key auth (single api_key parameter) but shows no support for scoped/least-privilege credentials such as role-based keys, granular permission scopes, or per-agent restricted tokens; only one flat API key model is documented.

   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “To start using Groq with OpenAI's client libraries, pass your Groq API key to the api_key parameter and change the base_url to https://api.g…”
   • [claimed-docs] “Spend Limits: Set automated spending limits and receive budget alerts”
   • [claimed-docs] “Spend Limits:** Set automated spending limits and receive budget alerts”

   proof ↗⚑ flag

   Together AI—none0/10

   No evidence of scoped or least-privilege API key issuance (e.g., per-project keys, role/permission scoping, or restricted-scope tokens for agents); docs only mention a single API key used for authentication, with no mention of scoping controls.

   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”

   proof ↗⚑ flag

8. ai-native userBuild against official SDKs

   weight 2 · round drawn
   Groq✓fullclaimed8/10

   Groq maintains an official TypeScript SDK (groq-typescript) with concrete usage examples (audio transcription, file uploads) and extensive first-party docs covering chat completions, streaming, tool use, structured outputs, and OpenAI-compatible client usage, all of which are core to building against official SDKs. Missing for 10: explicit evidence of an official Python SDK repo (only OpenAI-compatible base_url swap is shown for Python) and independent/community validation specifically of SDK reliability rather than general API performance.

   • [github] “Request parameters that correspond to file uploads can be passed in many different forms”
   • [github] “await client.audio.transcriptions.create({ model: 'whisper-large-v3-turbo', file: fs.createReadStream('/path/to/file'), });”
   • [github] “If you have access to Node fs we recommend using fs.createReadStream()... Or if you have the web File API you can pass a File instance”
   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “To start using Groq with OpenAI's client libraries, pass your Groq API key to the api_key parameter and change the base_url to https://api.g…”
   • [claimed-docs] “To enable streaming, set the parameter stream=True. The completion function will then return an iterator of completion deltas”
   • [claimed-docs] “Tool use (or function calling) is what transforms a language model from a conversational interface into an autonomous agent capable of takin…”
   • [claimed-docs] “Structured Outputs is a feature that ensures your model responses conform to your provided JSON Schema”

   proof ↗⚑ flag

   Together AI✓fullclaimed8/10

   Together AI provides an official TypeScript/JavaScript SDK (github repo) and OpenAI-compatible API/SDK usage, with docs covering chat, structured outputs, function calling, streaming, fine-tuning, and dedicated deployment all launchable via API/SDK. This directly supports AI-native, agentic development against official SDKs, including agent loop patterns and tool calling. Missing for 10: independent hands-on developer reviews specifically praising SDK quality/completeness, and no official Python SDK repo cited in evidence pack.

   • [github] “We provide support for streaming responses using Server Sent Events (SSE).”
   • [github] “This library provides convenient access to the Together REST API from server-side TypeScript or JavaScript.”
   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”
   • [claimed-docs] “Function calling (also called tool calling) lets LLMs respond with structured function names and arguments that you can execute in your appl…”
   • [claimed-docs] “To build agent loops, chain tool calls inside one response (multi-step), and conversations that thread tools across many turns (multi-turn).”
   • [claimed-docs] “You can launch a fine-tuning job from the console, through the API/SDK, or with the CLI”
   • [claimed-docs] “Supported models can return JSON that conforms to any schema you supply, so you can read the output directly in code without retries or frag…”

   proof ↗⚑ flag

9. ai-native userSubscribe to events via webhooks

   weight 2 · round drawn
   Groq—none0/10

   No evidence anywhere in the pack of a webhooks/event-subscription feature for Groq's API—only synchronous/streaming inference, batch, tool-use/MCP, and admin/billing docs are covered.

   ⚑ flag

   Together AI—none0/10

   No evidence of webhook subscription support anywhere in the docs, SDKs, or probes; Together AI offers SSE streaming and a docs MCP server, but nothing about webhook event subscriptions for async notifications (e.g., fine-tune job completion, batch job status).

   ⚑ flag

Agentic features

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

   weight 2 · round drawn
   Groq—none0/10

   Groq's evidence covers only inference primitives (chat completions, tool use, MCP tool integration, batch processing, structured outputs) — none of which describe a scheduling/trigger system for autonomous background automations; batch jobs are asynchronous request batches, not persistent background agents. Missing for 10: any documented scheduler, cron/trigger mechanism, persistent background agent runtime, or workflow orchestration feature that lets a user 'set up' an automation to run unattended.

   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”
   • [claimed-docs] “Tool use (or function calling) is what transforms a language model from a conversational interface into an autonomous agent capable of takin…”
   • [claimed-docs] “point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”

   proof ↗⚑ flag

   Together AI—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

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

   weight 3 · round drawn
   Groq—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

   Together AI—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

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

   weight 2 · round to Groq
   Groq~partialclaimed6/10

   Groq's Chat Completions/Responses API is inherently natural-language driven, and its tool-use and remote-MCP features let a model take autonomous actions from NL instructions, which supports agentic operation on top of the API. However, this is evidence of building NL-driven agents *with* Groq's models, not of an NL command interface for operating the Groq platform/product itself (e.g., managing keys, billing, deployments via chat). Missing for 10: an NL/chat interface for controlling the Groq platform itself, independent verification of agentic tool-use reliability in production.

   • [claimed-docs] “Generating text with Groq's Chat Completions API enables you to have natural, conversational interactions with Groq's large language models.”
   • [claimed-docs] “Tool use (or function calling) is what transforms a language model from a conversational interface into an autonomous agent capable of takin…”
   • [claimed-docs] “To use tools, the model must be provided with tool definitions. These tool definitions are in JSON schema format and are passed to the model…”
   • [claimed-docs] “point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”
   • [claimed-docs] “Groq's Responses API supports remote tool use via MCP servers via HTTPS where Groq handles all orchestration... You don't implement anything…”

   proof ↗⚑ flag

   Together AI—none0/10

   Together AI is an inference/training API platform operated via REST API, SDKs, CLI, and console — the evidence shows structured commands (API calls, CLI syntax like 'tg beta endpoints deploy...') rather than any natural-language command interface for operating the platform itself. The agent-skills/MCP docs only help external coding agents write correct Together AI code, not let a user operate Together AI's own product surface via natural language.

   ⚑ flag

Api quality

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

   weight 2 · round to Groq
   Groq~partialprobed4/10

   Groq's docs (console.groq.com/docs) are extensive and code-example-rich (curl/Python/JS snippets for chat, tool-use, structured outputs, batch, MCP, etc.), functioning as a de facto API reference, but there's no evidence of an interactive, in-browser 'runnable' playground or embedded code execution — probes show no discoverable OpenAPI/swagger spec (404s) and no mention of a try-it-now console feature. Missing for 10: an actual interactive/executable API reference (e.g., Swagger UI, embedded runnable code sandbox), confirmation of an OpenAPI spec, and independent user confirmation of using such an interactive reference.

   • [claimed-docs] “Fast LLM inference, OpenAI-compatible. Simple to integrate, easy to scale. Start building in minutes.”
   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “To enable streaming, set the parameter stream=True. The completion function will then return an iterator of completion deltas”
   • [probe] “PROBE docs-md: HTTP 200 at https://console.groq.com/docs/overview.md --- description: Fast LLM inference, OpenAI-compatible. Simple to integ…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://console.groq.com/openapi.json, https://console.groq.com/swagger.json, https://console.groq.c…”

   proof ↗⚑ flag

   Together AI—none0/10

   The evidence pack shows extensive prose documentation (guides, overview pages) but no reference to an interactive API reference or runnable/try-it-now code examples; a direct probe for OpenAPI/Swagger specs at standard paths returned 404 for all candidates, indicating no interactive spec-driven reference was found. No community or hands-on evidence confirms a runnable API console exists.

   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.together.ai/openapi.json, https://docs.together.ai/swagger.json, https://docs.together.…”
   • [probe] “PROBE llms.txt: HTTP 200 at https://docs.together.ai/llms.txt # Together AI docs > Documentation for the Together AI platform for inference…”
   • [probe] “PROBE docs-md: HTTP 200 at https://docs.together.ai/intro.md > ## Documentation Index > Fetch the complete documentation index at: https://d…”

   proof ↗⚑ flag

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

   weight 2 · round drawn
   Groq—none0/10

   Groq's API is OpenAI-compatible and well documented, but there is no evidence of a downloadable OpenAPI/Swagger spec; a direct probe of common spec paths (openapi.json, swagger.json, etc.) all returned 404, and no docs page links to a machine-readable spec.

   • [probe] “PROBE openapi: all candidate paths 404 (https://console.groq.com/openapi.json, https://console.groq.com/swagger.json, https://console.groq.c…”

   proof ↗⚑ flag

   Together AI—none0/10

   Active probing for standard OpenAPI/swagger spec locations (openapi.json, swagger.json, etc.) all returned 404, and no docs page in the pack references a downloadable OpenAPI/machine-readable spec for the Together API.

   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.together.ai/openapi.json, https://docs.together.ai/swagger.json, https://docs.together.…”

   proof ↗⚑ flag

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

   weight 1 · round drawn
   Groq—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

   Together AI—none0/10

   No evidence of a dedicated sandbox/staging environment; the only related mention is a test account with trial credit on the same production API, not an isolated sandbox for testing without touching production data.

   • [community] “the price is pretty low... 4B MODEL, PRICE 1K TOKENS: $0.0001. register with an email, test account has $25 credit, python API as well, good…”

   proof ↗⚑ flag

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

   weight 2 · round drawn
   Groq—none0/10

   No evidence pack item documents API versioning conventions or a formal deprecation policy for Groq's endpoints or models; docs cover OpenAI-compatibility, features, and pricing but not lifecycle/versioning guarantees.

   ⚑ flag

   Together AI—none0/10

   The evidence shows a documented, OpenAI-compatible REST API (with a /v1/ endpoint) but contains no mention of a versioning scheme or a documented deprecation policy for API changes. Absence of evidence for this applicable capability means it cannot be credited as delivered.

   ⚑ flag

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

   weight 2 · round drawn
   Groq✓fullclaimed7/10

   Groq's Batch API explicitly supports submitting thousands of API requests asynchronously at 50% lower cost with a 24h-7day processing window, directly satisfying bulk operations across many items, and higher-throughput Flex processing further supports large-scale workloads. Missing for 10: independent hands-on verification of batch job completion/reliability at scale and details on per-batch item limits or error handling.

   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”
   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”
   • [claimed-docs] “Flex processing is available for all models to paid customers only with 10x higher rate limits compared to on-demand processing.”
   • [claimed-docs] “Flex processing is available for all models to paid customers only with 10x higher rate limits compared to on-demand processing. Pricing mat…”

   proof ↗⚑ flag

   Together AI✓fullclaimed7/10

   Together AI offers an asynchronous Batch API explicitly designed for bulk workloads across many items at once, at up to 50% lower cost than synchronous calls, alongside dedicated endpoints and fine-tuning pipelines that can process large datasets end-to-end. missing for 10: no documented batch size/throughput limits, no independent/hands-on report validating batch job behavior at scale, and no detail on input format flexibility for large item counts.

   • [claimed-docs] “Run asynchronous batch workloads at up to 50% lower cost.”
   • [claimed-docs] “Together AI handles the full lifecycle: data upload, training, hosting, and inference on a dedicated endpoint.”
   • [claimed-docs] “Fine-tuning tailors a pretrained model to a smaller, targeted dataset so it performs better on a specific task or domain. Together AI handle…”

   proof ↗⚑ flag

Batch

1. ml-engineerSubmit asynchronous batch inference jobs at a documented discount versus real-time pricing

   weight 2 · round drawn
   Groq✓fullclaimed8/10

   Groq's official docs explicitly describe an asynchronous batch API delivering 50% lower cost than standard on-demand pricing, with a 24-hour to 7-day processing window and no impact on standard rate limits (groq-docs-7, groq-docs-29). This directly matches the story's requirement of documented discounted async batch inference. Missing for 10: independent/hands-on confirmation of the batch discount actually being applied in practice, and more detail on job submission/monitoring workflow beyond the docs summary.

   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”
   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”

   proof ↗⚑ flag

   Together AI✓fullclaimed8/10

   Together AI has documented batch inference at up to 50% lower cost than real-time serverless pricing, directly matching the story's requirement for asynchronous jobs at a documented discount [together-ai-docs-3]. missing for 10: independent/hands-on corroboration of actual batch job submission and discount realization beyond the docs page, and more detail on submission workflow/API specifics.

   • [claimed-docs] “Run asynchronous batch workloads at up to 50% lower cost.”

   proof ↗⚑ flag

Dedicated

1. ml-engineerDeploy a model on dedicated GPU capacity with autoscaling so my traffic is isolated from the shared serverless pool

   weight 2 · round to Together AI
   Groq—none0/10

   Evidence shows Groq's service tiers (on-demand, flex, auto) and batch processing, all within the shared serverless API, but no documentation of dedicated GPU capacity or customer-isolated infrastructure with autoscaling; one community remark about 'call us for quotes' for newer models hints at an enterprise track but gives no detail on dedicated capacity or autoscaling.

   • [claimed-docs] “Groq offers multiple service tiers so you can tune for latency, throughput, and reliability.”
   • [claimed-docs] “Flex processing is available for all models to paid customers only with 10x higher rate limits compared to on-demand processing.”
   • [claimed-docs] “auto: Pass this if you dont want to think about tiers and you want to leverage the best tier available to you at any given moment.”
   • [community] “As soon as i saw they switched to 'call us for quotes' for the new models, i knew they are over.”

   proof ↗⚑ flag

   Together AI✓fullclaimed8/10

   Docs explicitly describe Dedicated Model Inference (DMI) for serving models on reserved/isolated hardware distinct from the shared serverless pool, with no hard rate limits, and a dedicated scaling doc showing configurable autoscaling via replica limits, plus deployment of fine-tuned or custom models via CLI. This directly matches the story's requirements of dedicated GPU capacity, autoscaling, and isolation from shared serverless traffic. Missing for 10: independent/hands-on verification of autoscaling behavior in production and concrete GPU-type/SLA details beyond docs.

   • [claimed-docs] “Dedicated model inference (DMI) lets you serve a model on reserved hardware, providing several advantages over serverless models”
   • [claimed-docs] “Configure your deployment to scale automatically by setting limits on how many replicas it can run.”
   • [claimed-docs] “Fine-tuned models: Deploy a model you fine-tuned from a supported base model.”
   • [claimed-docs] “tg beta endpoints deploy google/gemma-4-E4B-it --endpoint my-endpoint”
   • [claimed-docs] “Dedicated model inference (DMI) lets you serve a model on reserved hardware... Better performance... No hard rate limits”

   proof ↗⚑ flag

Fine tune

1. ml-engineerFine-tune a supported base model on my own data and serve the result on the same platform

   weight 2 · round to Together AI
   Groq—none0/10

   Groq's docs show it supports serving pre-made LoRA adapters for inference (groq-docs-11, groq-docs-20, groq-docs-32), but there is no evidence of any fine-tuning/training capability on the platform itself — adapters must be trained elsewhere and uploaded. No mention of a fine-tuning API, training jobs, or dataset ingestion for model customization.

   • [claimed-docs] “Upload your existing LoRA adapters to run specialized inference while maintaining the performance and efficiency of Groq's infrastructure.”
   • [claimed-docs] “With LoRA inference on Groq, you can: Run inference with your pre-made LoRA adapters”
   • [claimed-docs] “Groq provides inference services for pre-made Low-Rank Adaptation (LoRA) adapters... Upload your existing LoRA adapters to run specialized i…”

   proof ↗⚑ flag

   Together AI✓fullclaimed9/10

   Docs clearly describe fine-tuning (LoRA or full) on supported base models via console/API/CLI, with Together AI handling the full lifecycle including data upload, training, hosting, and inference on a dedicated endpoint, and dedicated endpoints explicitly support deploying fine-tuned models. This is a first-party documented end-to-end workflow with concrete CLI examples. Missing for 10: independent/hands-on confirmation from a third party that a fine-tune-then-serve workflow works as described.

   • [claimed-docs] “Fine-tune models on your own data and deploy them for inference.”
   • [claimed-docs] “LoRA: Trains a small set of adapter weights on top of the frozen base model... Full fine-tuning: Updates every weight in the base model.”
   • [claimed-docs] “You can launch a fine-tuning job from the console, through the API/SDK, or with the CLI”
   • [claimed-docs] “Fine-tuned models: Deploy a model you fine-tuned from a supported base model.”
   • [claimed-docs] “Together AI handles the full lifecycle: data upload, training, hosting, and inference on a dedicated endpoint.”
   • [claimed-docs] “Fine-tuning tailors a pretrained model to a smaller, targeted dataset so it performs better on a specific task or domain. Together AI handle…”

   proof ↗⚑ flag

2. ml-engineerUpload and serve my own custom model weights or LoRA adapters

   weight 2 · round to Together AI
   Groq~partialclaimed6/10

   Groq's docs explicitly describe LoRA adapter upload and serving ('Upload your existing LoRA adapters to run specialized inference' — groq-docs-11/20/32), directly matching part of the story. However, evidence only covers LoRA adapters on top of existing base models, not uploading fully custom model weights, and there's no independent/hands-on confirmation of this LoRA workflow actually working in practice. Missing for 10: evidence of custom full-weight model upload/serving, and independent corroboration that LoRA upload works reliably in production.

   • [claimed-docs] “Upload your existing LoRA adapters to run specialized inference while maintaining the performance and efficiency of Groq's infrastructure.”
   • [claimed-docs] “With LoRA inference on Groq, you can: Run inference with your pre-made LoRA adapters”
   • [claimed-docs] “Groq provides inference services for pre-made Low-Rank Adaptation (LoRA) adapters... Upload your existing LoRA adapters to run specialized i…”

   proof ↗⚑ flag

   Together AI~partialclaimed7/10

   Together AI's docs clearly support fine-tuning (LoRA or full) and deploying the resulting fine-tuned models to dedicated inference endpoints, covering the LoRA-adapter serving use case end-to-end (data upload, training, hosting, inference). However, there's no evidence of uploading externally-trained/custom weights not produced via Together's own fine-tuning pipeline (e.g., bringing your own checkpoint from another framework). missing for 10: evidence of uploading externally-trained custom weights/checkpoints (not produced by Together's fine-tuning), and independent/hands-on confirmation of this workflow.

   • [claimed-docs] “LoRA: Trains a small set of adapter weights on top of the frozen base model... Full fine-tuning: Updates every weight in the base model.”
   • [claimed-docs] “Fine-tuned models: Deploy a model you fine-tuned from a supported base model.”
   • [claimed-docs] “Together AI handles the full lifecycle: data upload, training, hosting, and inference on a dedicated endpoint.”
   • [claimed-docs] “Fine-tuning tailors a pretrained model to a smaller, targeted dataset so it performs better on a specific task or domain. Together AI handle…”
   • [claimed-docs] “You can launch a fine-tuning job from the console, through the API/SDK, or with the CLI”

   proof ↗⚑ flag

Catalog

1. ml-engineerGet newly released open-weight models on the platform quickly after their public release

   weight 2 · round to Groq
   Groq~partialcommunity5/10

   Groq's docs list several open-weight models (e.g. GPT-OSS 120B, Llama, Mixtral) and community reports mention using Kimi K2 'fast' shortly after its release, suggesting some open-weight models are added promptly, but there is no explicit documentation of a release-to-availability SLA or a systematic track record of rapid onboarding. Missing for 10: explicit timeline commitments or a changelog showing new open-weight model releases added within days, and independent verification beyond a single anecdote.

   • [claimed-docs] “GPT-OSS 120B is OpenAI's flagship open-weight language model with 120 billion parameters, built in browser search and code execution, and re…”
   • [community] “I wanted to use Kimi K2 fast for coding and Groq was the only fast provider at the time... Definitely recommend cerebras tho now that groq's…”
   • [claimed-docs] “Upload your existing LoRA adapters to run specialized inference while maintaining the performance and efficiency of Groq's infrastructure.”

   proof ↗⚑ flag

   Together AI—none0/10

   The evidence pack shows Together AI hosts 100+ open-source models and supports fine-tuning/dedicated deployment, but there is no evidence—first-party or independent—about the speed or process of adding newly released open-weight models to the catalog. missing for 10: any documentation or community report on time-to-availability for new model releases, changelog/release-notes evidence, or comparison of catalog freshness versus competitors.

   ⚑ flag

2. developerRely on a documented deprecation policy with advance notice before a hosted model is removed

   weight 1 · round drawn
   Groq—none0/10

   No evidence pack item documents a deprecation policy, sunset timeline, or advance-notice commitment for hosted models being removed; community comments even reference abrupt shifts (e.g., new models moved to 'call us for quotes') without formal notice process. This is a fair question for a model-hosting API but no supporting documentation exists.

   • [community] “As soon as i saw they switched to 'call us for quotes' for the new models, i knew they are over.”

   proof ↗⚑ flag

   Together AI—none0/10

   No evidence in the pack mentions a deprecation policy, sunset schedule, or advance-notice commitment for retiring hosted models; docs cover serverless/dedicated inference, fine-tuning, and agent tooling but nothing about model lifecycle/deprecation notices.

   ⚑ flag

3. ai-native userHave an agent enumerate the live model catalog programmatically via a documented GET /v1/models-style endpoint

   weight 2 · round drawn
   Groq✓fullprobed8/10

   Groq exposes an OpenAI-compatible /v1/models endpoint (confirmed live via probe returning JSON error structure, not 404), and docs establish OpenAI SDK/client compatibility with base_url https://api.groq.com/openai/v1, which per OpenAI spec includes the models list endpoint; this is documented and independently verified via a real GET request. missing for 10: no explicit Groq-authored documentation page specifically describing the /v1/models endpoint response schema, and no successful authenticated enumeration example shown (only the 401 unauthenticated probe).

   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.groq.com/openai/v1/models without a key returned HTTP 401 ({"error":{"message":"Invalid …”
   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “To start using Groq with OpenAI's client libraries, pass your Groq API key to the api_key parameter and change the base_url to https://api.g…”
   • [claimed-docs] “Fast LLM inference, OpenAI-compatible. Simple to integrate, easy to scale. Start building in minutes.”

   proof ↗⚑ flag

   Together AI✓fullprobed8/10

   Together AI exposes an OpenAI-compatible /v1/models endpoint confirmed live via probe (returns 401 without key, meaning it exists and speaks JSON, per together-ai-probe-rt-1), and docs confirm OpenAI-compatibility for base URL/API key swaps (together-ai-docs-1) enabling programmatic catalog enumeration. missing for 10: a dedicated docs page explicitly documenting the /v1/models endpoint schema/response format, and independent hands-on confirmation of a successful authenticated call listing the full catalog.

   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.together.xyz/v1/models without a key returned HTTP 401 (Missing API key) — the OpenAI-st…”
   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”

   proof ↗⚑ flag

4. developerChoose among a broad catalog of current open-weight model families (Llama, Qwen, DeepSeek, GPT-OSS and peers) on shared serverless endpoints

   weight 3 · round to Together AI
   Groq~partialprobed6/10

   Groq's docs confirm OpenAI-compatible serverless endpoints hosting multiple model families, explicitly naming GPT-OSS 120B and (via community reports) Llama and Mixtral models, but the evidence pack never mentions Qwen or DeepSeek being hosted, so the 'broad catalog' claim is only partially substantiated. missing for 10: explicit documentation of Qwen and DeepSeek model availability, and a public model list/catalog page confirming breadth beyond GPT-OSS/Llama/Mixtral.

   • [claimed-docs] “GPT-OSS 120B is OpenAI's flagship open-weight language model with 120 billion parameters, built in browser search and code execution, and re…”
   • [claimed-docs] “To start using Groq with OpenAI's client libraries, pass your Groq API key to the api_key parameter and change the base_url to https://api.g…”
   • [community] “Incredible tool. The Mixtral 8x7B model running on their hardware did 491.40 T/s for me…”
   • [community] “I'm achieving consistent 450+ tokens/sec for Mixtral 8x7b 32k and ~200 tps for Llama 2 70B-4k.”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.groq.com/openai/v1/models without a key returned HTTP 401 ({"error":{"message":"Invalid …”

   proof ↗⚑ flag

   Together AI✓fullprobed7/10

   Docs confirm serverless endpoints serving '100+ open-source models' with per-token pricing, no provisioning latency, and OpenAI-compatible API for easy model switching, and community evidence corroborates real-world use of Llama-family models on the platform. However, the evidence pack never explicitly enumerates Qwen, DeepSeek, or GPT-OSS by name, so full breadth across all cited model families isn't directly confirmed. Missing for 10: explicit documentation naming Qwen/DeepSeek/GPT-OSS as catalog entries, and independent verification of catalog breadth beyond the generic '100+ models' claim.

   • [claimed-docs] “Call 100+ open-source models with per-token pricing and no provisioning latency.”
   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”
   • [claimed-docs] “you can prototype on serverless and move to reserved hardware later without changing your application code”
   • [community] “Colleague and I were discussing need for a service like this. Their pricing is much cheaper than self-hosting Llama-2-70B on AWS ($20.36/hr)…”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.together.xyz/v1/models without a key returned HTTP 401 (Missing API key) — the OpenAI-st…”

   proof ↗⚑ flag

Modalities

1. developerGenerate embeddings (and rerank results) for retrieval pipelines without a second vendor

   weight 1 · round to Together AI
   Groq—none0/10

   No evidence of an embeddings endpoint or reranking model/API in Groq's documentation; the evidence pack covers chat completions, tool use, transcription, LoRA, batch, and service tiers but never mentions embeddings or rerank capabilities.

   ⚑ flag

   Together AI~partialclaimed5/10

   Docs confirm a unified API that includes embeddings alongside chat/image/audio (together-ai-docs-21), supporting retrieval pipelines without a separate embeddings vendor. However, no evidence in the pack documents a dedicated rerank endpoint or reranking models, so the reranking half of the story is unsupported. Missing for 10: dedicated rerank API/model documentation, embeddings model list/dimensions, and independent confirmation of embeddings quality/usage in RAG pipelines.

   • [claimed-docs] “Call chat, image, audio, embedding, and more through one API.”

   proof ↗⚑ flag

2. developerCall vision, audio, or image-generation models beyond text chat on the same platform

   weight 1 · round to Together AI
   Groq~partialclaimed6/10

   Groq's docs confirm audio transcription (Whisper models via groq-gh-2) and vision input support through the Responses API (text and image inputs, groq-docs-13), extending beyond pure text chat. However, there is no evidence of image-generation models or text-to-speech/audio-generation capability on the platform. missing for 10: image-generation model support, audio-generation/TTS support, independent hands-on confirmation of multimodal calls.

   • [github] “await client.audio.transcriptions.create({ model: 'whisper-large-v3-turbo', file: fs.createReadStream('/path/to/file'), });”
   • [claimed-docs] “The Responses API supports both text and image inputs while producing text outputs, stateful conversations, and function calling to connect …”
   • [claimed-docs] “Groq's Responses API is fully compatible with OpenAI's Responses API, making it easy to integrate advanced conversational AI capabilities in…”

   proof ↗⚑ flag

   Together AI✓fullclaimed8/10

   Together AI's docs explicitly state you can 'Call chat, image, audio, embedding, and more through one API' on the serverless platform, indicating multimodal model support beyond text chat on the same platform [together-ai-docs-21]. This is corroborated by the broader serverless/dedicated inference docs describing a unified API/base URL approach [together-ai-docs-1, together-ai-docs-2]. Missing for 10: no independent hands-on report or example code specifically calling vision/audio/image-generation endpoints, and no detailed model list confirming breadth of non-text modalities.

   • [claimed-docs] “Call chat, image, audio, embedding, and more through one API.”
   • [claimed-docs] “Call 100+ open-source models with per-token pricing and no provisioning latency.”
   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”

   proof ↗⚑ flag

Compat

1. ai-native userPlug the provider into coding agents and agent frameworks through documented, first-party integration guides

   weight 2 · round to Together AI
   Groq~partialprobed7/10

   Groq's docs clearly document OpenAI-compatible base_url swapping (groq-docs-2/25), tool-use/function-calling (groq-docs-5/27), a Responses API that is 'fully compatible with OpenAI's Responses API' (groq-docs-23), and first-party remote MCP server support where 'Groq handles all orchestration' (groq-docs-6/17/28, groq-probe-4) — all of which make it straightforward to plug into agent frameworks that speak the OpenAI protocol or MCP. However, there are no named first-party guides for specific coding-agent/agent-framework integrations (e.g., LangChain, Cursor, AutoGPT) beyond the generic OpenAI-compat pointer, and community reports flag real-world reliability quirks (groq-comm-12/13/14) that add friction beyond the docs' 'simple to integrate' framing. Missing for 10: named integration guides for specific agent frameworks/coding agents, and independent hands-on confirmation of smooth agent-framework plug-in.

   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “To start using Groq with OpenAI's client libraries, pass your Groq API key to the api_key parameter and change the base_url to https://api.g…”
   • [claimed-docs] “Tool use (or function calling) is what transforms a language model from a conversational interface into an autonomous agent capable of takin…”
   • [claimed-docs] “To use tools, the model must be provided with tool definitions. These tool definitions are in JSON schema format and are passed to the model…”
   • [claimed-docs] “point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”
   • [claimed-docs] “you simply point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”
   • [claimed-docs] “Groq's Responses API supports remote tool use via MCP servers via HTTPS where Groq handles all orchestration... You don't implement anything…”
   • [claimed-docs] “Groq's Responses API is fully compatible with OpenAI's Responses API, making it easy to integrate advanced conversational AI capabilities in…”
   • [probe] “official MCP server documented at https://console.groq.com/docs/mcp”
   • [community] “My company had a really terrible experience trying to use Groq, and I would NOT recommend anyone use their service if you need reliability. …”
   • [community] “I don't really get the value proposition of groq as a user, the performance is really poor for the token price.”

   proof ↗⚑ flag

   Together AI✓fullprobed9/10

   Together AI documents first-party agent-skills and an MCP server specifically for coding agent integration (docs-agent-skills, docs-12/19/20/24/29/30), confirmed live via probe (probe-4, probe-rt-3), plus OpenAI-compatible base URL swap for framework compatibility (docs-1) and function/tool-calling docs for agentic workflows (docs-8/9). Missing for 10: independent hands-on reports specifically validating the agent-skills/MCP integration with popular frameworks like LangChain or Claude Code.

   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”
   • [claimed-docs] “Make your AI coding agent Together-AI-aware with ready-made skills for code generation and an MCP server for live docs lookup.”
   • [claimed-docs] “12 domain-specific skills that load on demand and teach your agent how to write correct Together AI code (right model IDs, SDK patterns, bes…”
   • [claimed-docs] “Gives your agent live access to this documentation site so it can look up current information without leaving your editor.”
   • [claimed-docs] “Docs MCP server: Gives your agent live access to this documentation site so it can look up current information without leaving your editor.”
   • [claimed-docs] “Function calling (also called tool calling) lets LLMs respond with structured function names and arguments that you can execute in your appl…”
   • [claimed-docs] “To build agent loops, chain tool calls inside one response (multi-step), and conversations that thread tools across many turns (multi-turn).”
   • [probe] “official MCP server documented at https://docs.together.ai/docs/agent-skills”
   • [probe] “PROBE mcp-endpoint (2026-09-04): POST initialize to https://docs.together.ai/mcp answered HTTP 200 with a JSON-RPC/MCP response (event: mess…”

   proof ↗⚑ flag

2. ai-native userHave an agent switch to or away from this provider mid-workflow because it speaks the standard chat-completions API without provider-specific code changes

   weight 3 · round drawn
   Groq✓fullprobed8/10

   Groq's docs explicitly instruct developers to reuse OpenAI client libraries by just swapping api_key and base_url to https://api.groq.com/openai/v1, and this is corroborated by a live probe showing the OpenAI-style /v1/models endpoint responding in the expected JSON error format. Chat completions, streaming, tool-use, and structured outputs are all documented against this same OpenAI-compatible surface, supporting drop-in provider switching mid-workflow. Missing for 10: independent/hands-on confirmation from a real agent framework that switching providers requires zero code changes, and clarity on any OpenAI parameters/features that are NOT supported (full parity claim not verified end-to-end).

   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “To start using Groq with OpenAI's client libraries, pass your Groq API key to the api_key parameter and change the base_url to https://api.g…”
   • [claimed-docs] “Fast LLM inference, OpenAI-compatible. Simple to integrate, easy to scale. Start building in minutes.”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.groq.com/openai/v1/models without a key returned HTTP 401 ({"error":{"message":"Invalid …”
   • [claimed-docs] “Generating text with Groq's Chat Completions API enables you to have natural, conversational interactions with Groq's large language models.”

   proof ↗⚑ flag

   Together AI✓fullprobed8/10

   Docs explicitly state OpenAI-compatibility requires only changing API key and base URL, meaning agents/tools written against the standard chat-completions API can point at Together with no provider-specific code changes; the live /v1/models endpoint confirms OpenAI-style REST shape. This directly supports switching to/away from Together mid-workflow with minimal friction. Missing for 10: independent third-party confirmation of drop-in swap behavior (e.g., a framework's changelog showing zero-code-change provider swap) and explicit mention of streaming/tool-call parity with OpenAI's spec beyond docs.

   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”
   • [claimed-docs] “you can prototype on serverless and move to reserved hardware later without changing your application code”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.together.xyz/v1/models without a key returned HTTP 401 (Missing API key) — the OpenAI-st…”

   proof ↗⚑ flag

3. developerPoint an existing OpenAI SDK client at the provider by changing only the base URL and API key

   weight 3 · round to Groq
   Groq✓fullprobed9/10

   Groq's docs explicitly instruct developers to point OpenAI SDK clients at Groq by only changing api_key and base_url to https://api.groq.com/openai/v1, and a live probe confirms the OpenAI-style /v1/models endpoint is functional. This is corroborated by consistent documentation across chat completions, streaming, tool use, and structured outputs all using OpenAI SDK conventions. Missing for 10: independent third-party confirmation of a drop-in swap working flawlessly in production (community evidence focuses on speed/reliability rather than compatibility specifics).

   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “To start using Groq with OpenAI's client libraries, pass your Groq API key to the api_key parameter and change the base_url to https://api.g…”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.groq.com/openai/v1/models without a key returned HTTP 401 ({"error":{"message":"Invalid …”
   • [claimed-docs] “Fast LLM inference, OpenAI-compatible. Simple to integrate, easy to scale. Start building in minutes.”

   proof ↗⚑ flag

   Together AI✓fullprobed8/10

   Docs explicitly state you can point an existing client at Together by changing only the API key and base URL, and the API is OpenAI-compatible with a live /v1/models endpoint confirmed via probe. missing for 10: independent developer testimonial specifically confirming drop-in OpenAI SDK compatibility (community evidence covers pricing/UX but not this specific mechanic), and no explicit code snippet shown in evidence.

   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.together.xyz/v1/models without a key returned HTTP 401 (Missing API key) — the OpenAI-st…”
   • [claimed-docs] “you can prototype on serverless and move to reserved hardware later without changing your application code”

   proof ↗⚑ flag

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

   weight 2 · round to Together AI
   Groq~partialprobed6/10

   Groq's API surface covers essentially all the model-facing capabilities shown in its console/docs — chat completions, streaming, tool use, structured outputs, batch processing, service tiers, LoRA inference, and MCP remote tool use are all documented as API features (groq-docs-2,5,6,7,8,11,16,28). However, account/billing controls like spend limits (groq-docs-12/24) appear to be console-only with no documented API equivalent, and the probe found no public OpenAPI spec (all candidate paths 404, groq-probe-3), making it unclear whether the full API surface is discoverable/self-describing the way the UI is. Missing for 10: API-based management of billing/spend limits, a published OpenAPI/schema confirming full UI-API parity, and independent confirmation that every console feature (e.g., playground-only settings) has an API equivalent.

   • [claimed-docs] “pass your Groq API key to the `api_key` parameter and change the `base_url` to `https://api.groq.com/openai/v1`”
   • [claimed-docs] “Tool use (or function calling) is what transforms a language model from a conversational interface into an autonomous agent capable of takin…”
   • [claimed-docs] “point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”
   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”
   • [claimed-docs] “Upload your existing LoRA adapters to run specialized inference while maintaining the performance and efficiency of Groq's infrastructure.”
   • [claimed-docs] “Structured Outputs is a feature that ensures your model responses conform to your provided JSON Schema”
   • [claimed-docs] “Groq's Responses API supports remote tool use via MCP servers via HTTPS where Groq handles all orchestration... You don't implement anything…”
   • [claimed-docs] “Spend Limits: Set automated spending limits and receive budget alerts”
   • [probe] “PROBE openapi: all candidate paths 404 (https://console.groq.com/openapi.json, https://console.groq.com/swagger.json, https://console.groq.c…”

   proof ↗⚑ flag

   Together AI~partialprobed7/10

   Docs show broad API/CLI/SDK coverage mirroring UI actions—fine-tuning can be launched 'from the console, through the API/SDK, or with the CLI', dedicated endpoints can be deployed via CLI (tg beta endpoints deploy), and inference, batch, and cluster provisioning are all API-accessible—suggesting strong UI/API parity for core workflows. However, there's no explicit documentation enumerating full parity (e.g., billing, account/org management, usage dashboards) or confirming every UI-only feature has an API equivalent, and no OpenAPI spec was found (probe shows all openapi.json paths 404), making completeness hard to verify. Missing for 10: an explicit parity statement or API coverage for account/billing/admin functions, and a published OpenAPI spec confirming full API surface.

   • [claimed-docs] “You can launch a fine-tuning job from the console, through the API/SDK, or with the CLI”
   • [claimed-docs] “tg beta endpoints deploy google/gemma-4-E4B-it --endpoint my-endpoint”
   • [claimed-docs] “Together AI handles the full lifecycle: data upload, training, hosting, and inference on a dedicated endpoint.”
   • [claimed-docs] “Spin up H100 and B200 clusters with attached storage for training or large batch jobs.”
   • [claimed-docs] “Run asynchronous batch workloads at up to 50% lower cost.”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.together.ai/openapi.json, https://docs.together.ai/swagger.json, https://docs.together.…”

   proof ↗⚑ flag

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

   weight 3 · round drawn
   Groq—none0/10

   Groq's docs cover API usage, models, tool-use, batch processing and billing, but nothing addresses exporting account data, conversation history, or batch outputs in open/portable formats, nor any account-closure data dump. Even though Groq's Responses API is 'stateful,' there is no documented export or data-portability mechanism.

   ⚑ flag

   Together AI—none0/10

   The evidence pack covers inference, fine-tuning, dedicated endpoints, and agent tooling, but contains no mention of a data export feature, downloadable account data, or a documented way to retrieve fine-tuning datasets/model weights in open formats and leave the platform. This is a fair question for a hosted AI platform that stores prompts, fine-tuning data, and trained models, so absence of evidence means 'none' rather than 'na'.

   ⚑ flag

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

   weight 2 · round drawn
   Groq—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

   Together AI—none0/10

   The evidence pack covers Together AI's hosted inference, fine-tuning, and dedicated endpoint features, and its client SDKs (e.g., together-typescript) are open on GitHub, but there is no evidence that Together AI's core platform/product source is released under an open license—only that it serves open-source models to customers.

   ⚑ flag

Limits

1. developerRead documented rate limits and how they scale across usage tiers before I hit them in production

   weight 2 · round to Together AI
   Groq~partialclaimed5/10

   Groq's docs describe multiple service tiers (on-demand, flex, auto) and note Flex Processing gives '10x higher rate limits' for paid customers, plus batch processing having 'no impact to your standard rate limits' and spend-limit alerts — showing tiered scaling exists and is documented at a high level. However, no evidence pack citation shows a dedicated rate-limits page with concrete numeric request/token-per-minute limits per tier (free vs paid vs flex), so a developer can't get exact production-planning numbers from what's cited. Missing for 10: a documented numeric rate-limit table per usage tier, and independent confirmation that these documented limits match real-world behavior.

   • [claimed-docs] “Groq offers multiple service tiers so you can tune for latency, throughput, and reliability.”
   • [claimed-docs] “Flex processing is available for all models to paid customers only with 10x higher rate limits compared to on-demand processing.”
   • [claimed-docs] “Flex Processing is a service tier optimized for high-throughput workloads that prioritizes fast inference and can handle occasional request …”
   • [claimed-docs] “auto: Pass this if you dont want to think about tiers and you want to leverage the best tier available to you at any given moment.”
   • [claimed-docs] “Batch processing lets you run thousands of API requests at scale by submitting your workload as an asynchronous batch of requests to Groq wi…”
   • [claimed-docs] “Flex processing is available for all models to paid customers only with 10x higher rate limits compared to on-demand processing. Pricing mat…”
   • [claimed-docs] “Groq offers multiple service tiers so you can tune for latency, throughput, and reliability. You can distinguish these by providing the serv…”
   • [claimed-docs] “Spend Limits:** Set automated spending limits and receive budget alerts”

   proof ↗⚑ flag

   Together AI~partialclaimed6/10

   Together AI documents a rate-limit model explicitly (docs-14): dynamic per-organization, per-model limits based on live model capacity and recent successful usage, and notes dedicated endpoints have no hard rate limits (docs-31). This gives developers some visibility into how limits work before hitting them, but there's no documented tiered structure (e.g., spend-based tiers with specific numeric thresholds) showing how limits scale as usage/spend grows over time. Missing for 10: explicit usage-tier table or thresholds, numeric rate-limit values per tier, guidance on how to request/track tier upgrades.

   • [claimed-docs] “Together uses dynamic rate limits instead of fixed thresholds. Each organization has a dynamic rate per model that adjusts based on: The mod…”
   • [claimed-docs] “Dedicated model inference (DMI) lets you serve a model on reserved hardware... Better performance... No hard rate limits”

   proof ↗⚑ flag

Pricing

1. founderSet spending caps or budget alerts so a runaway workload cannot generate an unbounded bill

   weight 1 · round to Groq
   Groq✓fullclaimed8/10

   Groq's billing docs explicitly state users can set automated spending limits and receive budget alerts, directly matching the story. Missing for 10: no independent/hands-on corroboration of the spend-limit UI actually working, and no detail on granularity (per-project vs account-wide) or alert channels.

   • [claimed-docs] “Spend Limits: Set automated spending limits and receive budget alerts”
   • [claimed-docs] “Spend Limits:** Set automated spending limits and receive budget alerts”

   proof ↗⚑ flag

   Together AI—none0/10

   No evidence of spending caps, budget alerts, or usage-based cutoffs; docs mention rate limits (dynamic, based on capacity/usage) and pricing structures but nothing about setting a spend ceiling or receiving alerts to prevent runaway bills.

   ⚑ flag

2. founderSee public per-token prices for every hosted model without talking to sales

   weight 3 · round to Together AI
   Groq!disputedcontradicted4/10

   Groq's docs discuss billing mechanics (spend limits, flex-processing pricing parity, batch discount) implying a public pricing model, but no evidence pack item shows an actual public per-token price sheet for every hosted model, and a community report explicitly states Groq switched to 'call us for quotes' for newer models, contradicting the claim of fully public pricing without sales contact. Missing for 10: an actual public pricing page listing per-token rates for each model, and confirmation that newer/premium models are not gated behind sales quotes.

   • [claimed-docs] “Spend Limits: Set automated spending limits and receive budget alerts”
   • [claimed-docs] “Spend Limits:** Set automated spending limits and receive budget alerts”
   • [claimed-docs] “Flex processing is available for all models to paid customers only with 10x higher rate limits compared to on-demand processing. Pricing mat…”
   • [community] “As soon as i saw they switched to 'call us for quotes' for the new models, i knew they are over.”
   • [community] “Very impressive! I am even more impressed by the API pricing though - 0.27/1M tokens seems like an order of magnitude cheaper than the GPT-3…”

   proof ↗⚑ flag

   Together AI~partialprobed6/10

   Docs advertise "per-token pricing" for 100+ serverless models and community evidence confirms self-serve signup with visible per-token cost (e.g. "$0.0001/1K tokens", $25 trial credit) with no sales contact required. However, no evidence pack item directly shows a public pricing page enumerating rates for every hosted model, and the live models API requires an API key to enumerate the catalog. Missing for 10: a documented public pricing page/table covering all models, independent confirmation that all model prices (not just examples) are visible pre-signup.

   • [claimed-docs] “Call 100+ open-source models with per-token pricing and no provisioning latency.”
   • [community] “the price is pretty low... 4B MODEL, PRICE 1K TOKENS: $0.0001. register with an email, test account has $25 credit, python API as well, good…”
   • [community] “I've explored most of the inference as a service platforms, theirs is particularly nice. (Clear/simple UI, good pricing, good speed.)”
   • [probe] “PROBE models-endpoint (2026-09-04): GET https://api.together.xyz/v1/models without a key returned HTTP 401 (Missing API key) — the OpenAI-st…”

   proof ↗⚑ flag

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

   weight 2 · round drawn
   Groq—none0/10

   No evidence in the pack mentions data residency, region selection, or geographic storage controls anywhere in Groq's docs, GitHub, or community discussion.

   ⚑ flag

   Together AI—none0/10

   No evidence in the pack mentions data residency, regional storage options, or geographic control over where data/models are hosted; only privacy policy criticism unrelated to residency is present. Missing for 10: any documentation of region selection, data residency guarantees, or geo-specific hosting options.

   ⚑ flag

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

   weight 3 · round drawn
   Groq—none0/10

   No evidence pack item addresses data usage/training policies, opt-out controls, or privacy commitments regarding whether user data is used to train Groq's models.

   ⚑ flag

   Together AI—none0/10

   No first-party documentation describes an opt-out, data-retention control, or 'no training on your data' policy; the only relevant evidence is a community report describing Together AI's privacy policy as broad/vague, allowing data use 'for other purposes,' which points toward the opposite of a training-opt-out guarantee.

   • [community] “Together AI's privacy policy collects broad personal data including prompts, IP addresses, and contact info, with vague language allowing us…”
   • [community] “Colleague and I were discussing need for a service like this. Their pricing is much cheaper than self-hosting Llama-2-70B on AWS ($20.36/hr)…”

   proof ↗⚑ flag

3. ai-native userControl data retention and deletion

   weight 2 · round drawn
   Groq—none0/10

   No evidence pack items address data retention policies, deletion controls, or user-facing privacy/data management settings for Groq's API; the docs focus on inference features, tooling, pricing, and rate limits with nothing on retention/deletion.

   ⚑ flag

   Together AI—none0/10

   No evidence pack item documents user-facing data retention or deletion controls (e.g., API/dashboard settings to delete stored prompts or data). The only relevant community evidence (together-ai-comm-1) actually raises concerns that Together AI's privacy policy allows broad data collection with vague retention/usage language, which is the opposite of demonstrating retention/deletion controls.

   • [community] “Together AI's privacy policy collects broad personal data including prompts, IP addresses, and contact info, with vague language allowing us…”

   proof ↗⚑ flag

4. ai-native userOpt out of telemetry and usage tracking

   weight 2 · round drawn
   Groq—none0/10

   No evidence pack item addresses telemetry, usage tracking, opt-out settings, or privacy controls for Groq's API or console; the docs cover inference features, pricing, and infra but nothing about data collection preferences.

   ⚑ flag

   Together AI—none0/10

   No documentation or feature reference describes a telemetry/usage-tracking opt-out mechanism; the only related evidence is a community critique noting Together AI's privacy policy broadly collects prompts, IPs, and contact info with vague 'other purposes' language and no opt-out mentioned, reinforcing the absence of this capability.

   • [community] “Together AI's privacy policy collects broad personal data including prompts, IP addresses, and contact info, with vague language allowing us…”
   • [community] “Colleague and I were discussing need for a service like this. Their pricing is much cheaper than self-hosting Llama-2-70B on AWS ($20.36/hr)…”

   proof ↗⚑ flag

Reliability

1. founderGet a stated availability SLA on paid or enterprise tiers

   weight 1 · round drawn
   Groq—none0/10

   Evidence shows service tiers, flex processing, and a public status page, but nowhere is there a stated uptime/response-time SLA commitment for paid or enterprise customers; enterprise tiers are only referenced via 'call us for quotes' community comments with no SLA terms disclosed.

   • [claimed-docs] “Groq offers multiple service tiers so you can tune for latency, throughput, and reliability.”
   • [claimed-docs] “Groq offers multiple service tiers so you can tune for latency, throughput, and reliability. You can distinguish these by providing the serv…”
   • [probe] “PROBE status-page (2026-09-04): https://groqstatus.com returns HTTP 200 and renders a public service-status page (page body includes "operat…”
   • [community] “As soon as i saw they switched to 'call us for quotes' for the new models, i knew they are over.”

   proof ↗⚑ flag

   Together AI—none0/10

   Evidence shows a public status page and various product docs, but nothing states an actual uptime/availability SLA percentage or enterprise contractual guarantee. Absence of evidence for this applicable capability yields none.

   • [probe] “PROBE status-page (2026-09-04): https://status.together.ai returns HTTP 200 and renders a public service-status page (page body includes "op…”

   proof ↗⚑ flag

2. founderCheck a public status page with incident history before betting production traffic on the platform

   weight 2 · round to Together AI
   Groq~partialprobed5/10

   A probe confirms a live public status page at groqstatus.com that returns HTTP 200 and shows 'operational' status, indicating Groq does operate a status page a founder could check before committing production traffic. However, the evidence doesn't confirm the page includes detailed incident history, uptime history, or granular component-level status, and there's no first-party documentation describing the status page's features. missing for 10: confirmation of incident history/log on the status page, documented SLA or uptime commitments, and independent corroboration beyond a single automated probe.

   • [probe] “PROBE status-page (2026-09-04): https://groqstatus.com returns HTTP 200 and renders a public service-status page (page body includes "operat…”

   proof ↗⚑ flag

   Together AI✓fullprobed8/10

   A live public status page at status.together.ai returns HTTP 200 and shows service status ('operational'), confirming a real status dashboard exists that a founder could check before committing production traffic. Missing for 10: no evidence of detailed incident history/postmortems or uptime SLA history visible in the pack, so depth of incident history can't be fully confirmed.

   • [probe] “PROBE status-page (2026-09-04): https://status.together.ai returns HTTP 200 and renders a public service-status page (page body includes "op…”

   proof ↗⚑ flag

Benchmarks

1. ml-engineerSee published tokens-per-second or latency numbers, benchmarks, or load-testing guides backing the provider's speed claims

   weight 2 · round to Groq
   Groq!disputedcontradicted5/10

   There is no first-party published benchmark or load-testing page in the evidence (docs cover service tiers, flex processing, batch, streaming, but no concrete TPS/latency numbers or methodology). Independent hands-on community reports do supply concrete numbers (491 T/s and 450+ TPS for Mixtral, deterministic low-latency architecture claims) supporting the speed claim, but later community threads directly contradict this, calling performance 'really poor for the token price' and reliability 'worse than traditional hosts' with frequent errors. Missing for 10: an official Groq benchmark/load-testing page with methodology, and resolution of the conflicting recent community performance complaints.

   • [community] “Incredible tool. The Mixtral 8x7B model running on their hardware did 491.40 T/s for me…”
   • [community] “I'm achieving consistent 450+ tokens/sec for Mixtral 8x7b 32k and ~200 tps for Llama 2 70B-4k.”
   • [community] “Groq AMA: 'Unlike with graphics processors, which really need data parallelism to get good throughput, our LPU architecture allows us to del…”
   • [community] “Groq staff: our system is deterministic, no need for waiting or queuing anywhere, and we can have very low latency interconnect between card…”
   • [community] “There's a trail of complaints going back years now, and they rounded out the bottom of Kimi's verification program. Groq hosted models were/…”
   • [community] “I don't really get the value proposition of groq as a user, the performance is really poor for the token price.”
   • [claimed-docs] “Groq offers multiple service tiers so you can tune for latency, throughput, and reliability.”
   • [claimed-docs] “Flex processing is available for all models to paid customers only with 10x higher rate limits compared to on-demand processing.”

   proof ↗⚑ flag

   Together AI—none0/10

   The evidence pack contains only vague qualitative claims ('no provisioning latency', 'Better performance' for dedicated endpoints) and subjective community remarks ('good speed') but no published tokens-per-second figures, latency benchmarks, or load-testing methodology/guides that would let an ml-engineer verify speed claims.

   • [claimed-docs] “Call 100+ open-source models with per-token pricing and no provisioning latency.”
   • [claimed-docs] “Dedicated model inference (DMI) lets you serve a model on reserved hardware... Better performance... No hard rate limits”
   • [community] “I've explored most of the inference as a service platforms, theirs is particularly nice. (Clear/simple UI, good pricing, good speed.)”

   proof ↗⚑ flag

Serving

1. developerServe latency-sensitive workloads with fast time-to-first-token and high-throughput generation

   weight 3 · round to Groq
   Groq!disputedcontradicted6/10

   Early hands-on community reports strongly corroborate Groq's speed claims (491 T/s and ~450 t/s on Mixtral, deterministic low-latency architecture, streaming support docs), and Groq documents service tiers/flex-processing explicitly for latency/throughput tuning. However, more recent independent reports directly contradict this, describing Groq's hosted model performance as 'really poor for the token price' and 'always worse than traditional hosts,' plus reliability complaints ('so many random errors'), which concretely undercuts the fast/high-throughput positioning for current models. Missing for 10: recent (not 2024-era) throughput benchmarks, first-party TTFT numbers, and resolution of the conflicting recent user reports on performance-for-price.

   • [community] “Incredible tool. The Mixtral 8x7B model running on their hardware did 491.40 T/s for me…”
   • [community] “I'm achieving consistent 450+ tokens/sec for Mixtral 8x7b 32k and ~200 tps for Llama 2 70B-4k.”
   • [community] “Groq AMA: 'Unlike with graphics processors, which really need data parallelism to get good throughput, our LPU architecture allows us to del…”
   • [community] “Groq staff: our system is deterministic, no need for waiting or queuing anywhere, and we can have very low latency interconnect between card…”
   • [claimed-docs] “Groq offers multiple service tiers so you can tune for latency, throughput, and reliability.”
   • [claimed-docs] “Flex Processing is a service tier optimized for high-throughput workloads that prioritizes fast inference and can handle occasional request …”
   • [claimed-docs] “To enable streaming, set the parameter stream=True. The completion function will then return an iterator of completion deltas”
   • [community] “My company had a really terrible experience trying to use Groq, and I would NOT recommend anyone use their service if you need reliability. …”
   • [community] “There's a trail of complaints going back years now, and they rounded out the bottom of Kimi's verification program. Groq hosted models were/…”
   • [community] “I don't really get the value proposition of groq as a user, the performance is really poor for the token price.”

   proof ↗⚑ flag

   Together AI!disputedcontradicted5/10

   Together AI's docs claim strong speed characteristics — 'no provisioning latency' for serverless, DMI offering 'better performance' and 'no hard rate limits', dynamic rate limits based on live capacity, streaming SSE support, and automatic prompt caching to speed repeated prefills — all directly supporting fast TTFT and high throughput [together-ai-docs-2][together-ai-docs-31][together-ai-docs-14][together-ai-gh-1][together-ai-docs-13]. However, an independent hands-on comparison explicitly disputes this, stating Together AI was 'at the bottom of the pile' for large-scale deployment speed/pricing versus competitors like Cerebras, DeepInfra, and Hyperbolic [together-ai-comm-5], while another community post gives a positive but vaguer 'good speed' impression [together-ai-comm-4] — a genuine mixed/contradicted signal rather than mere skepticism. Missing for 10: independent benchmarks (TTFT/tokens-per-second numbers) confirming or refuting the vendor's latency claims, and resolution of the conflicting community reports.

   • [claimed-docs] “Call 100+ open-source models with per-token pricing and no provisioning latency.”
   • [claimed-docs] “Dedicated model inference (DMI) lets you serve a model on reserved hardware... Better performance... No hard rate limits”
   • [claimed-docs] “Together uses dynamic rate limits instead of fixed thresholds. Each organization has a dynamic rate per model that adjusts based on: The mod…”
   • [claimed-docs] “Automatic: There is no header, parameter, or account toggle to enable it. Send the same prompt prefix again and any portion that's still war…”
   • [github] “We provide support for streaming responses using Server Sent Events (SSE).”
   • [community] “I've explored most of the inference as a service platforms, theirs is particularly nice. (Clear/simple UI, good pricing, good speed.)”
   • [community] “'Together AI has quickly grown into the leading hosting platform for AI' - Lol. Definitely not. I went through pricing and large scale deplo…”

   proof ↗⚑ flag

2. ml-engineerBenefit from prompt/prefix caching that reduces latency or cost on repeated context

   weight 1 · round to Together AI
   Groq—none0/10

   No evidence pack item mentions prompt/prefix caching, context caching, or any discounted repeated-context pricing on Groq's API. Docs cover streaming, batch processing, service tiers, and flex processing, but none describe caching repeated prompt prefixes to cut latency or cost, so this applicable axis is unaddressed.

   ⚑ flag

   Together AI✓fullclaimed8/10

   Together AI documents automatic prompt/prefix caching for serverless chat models, explicitly stating cached input tokens are billed at a steep discount with no toggle needed, directly reducing cost on repeated context. missing for 10: independent/hands-on benchmark confirming latency reduction, and clarity on which models/latency gains apply beyond the cost-discount framing.

   • [claimed-docs] “Automatic: There is no header, parameter, or account toggle to enable it. Send the same prompt prefix again and any portion that's still war…”
   • [claimed-docs] “Send the same prompt prefix again and any portion that's still warm in the shared cache is billed at the cached rate.”
   • [claimed-docs] “Select serverless chat models bill cached input tokens at a steep discount... Automatic: There is no header, parameter, or account toggle to…”

   proof ↗⚑ flag

3. developerStream completions token by token over SSE for responsive user experiences

   weight 3 · round to Groq
   Groq✓fullcommunity9/10

   Docs explicitly document setting stream=True for token-by-token delta streaming via the OpenAI-compatible API, aligned with real-time display use cases, and independent community testimonials corroborate very high token throughput consistent with responsive streaming UX. Missing for 10: explicit confirmation of SSE transport format (docs describe iterator/deltas but don't spell out 'SSE' verbatim) and no independent hands-on benchmark specifically of streaming latency/SSE behavior.

   • [claimed-docs] “To enable streaming, set the parameter stream=True. The completion function will then return an iterator of completion deltas”
   • [claimed-docs] “To enable streaming, set the parameter `stream=True`.”
   • [claimed-docs] “you can stream the model's response in real-time. This allows your application to display the response as it's being generated”
   • [claimed-docs] “To enable streaming, set the parameter stream=True. The completion function will then return an iterator of completion deltas rather than a …”
   • [community] “Incredible tool. The Mixtral 8x7B model running on their hardware did 491.40 T/s for me…”
   • [community] “I'm achieving consistent 450+ tokens/sec for Mixtral 8x7b 32k and ~200 tps for Llama 2 70B-4k.”
   • [community] “Very impressed with the speed. This is one of the most impressive tech demos I've ever seen in my life... surreal to see the thing spitting …”

   proof ↗⚑ flag

   Together AI✓fullclaimed8/10

   Together AI's SDKs explicitly document SSE-based streaming support ('We provide support for streaming responses using Server Sent Events (SSE)'), and the API is OpenAI-compatible, implying standard streaming chat completion endpoints. Missing for 10: no independent hands-on latency benchmark or explicit token-by-token example snippet in the evidence pack.

   • [github] “We provide support for streaming responses using Server Sent Events (SSE).”
   • [github] “This library provides convenient access to the Together REST API from server-side TypeScript or JavaScript.”
   • [claimed-docs] “you can point it at models hosted on Together with two changes: the API key and base URL”

   proof ↗⚑ flag

Structured

1. developerEnforce structured outputs against a JSON schema (or grammar) so model responses parse reliably

   weight 3 · round drawn
   Groq✓fullclaimed8/10

   Groq's docs explicitly describe Structured Outputs with JSON Schema and a strict:true mode that uses constrained decoding to guarantee schema-conforming output, plus tool-use docs confirming JSON schema-based function calling. This is first-party documented and consistent with OpenAI-compatible API design. Missing for 10: independent/hands-on verification that strict mode reliably parses in production and no community corroboration of structured-output reliability specifically.

   • [claimed-docs] “With strict: true, the model uses constrained decoding to guarantee that the output will always match your schema exactly.”
   • [claimed-docs] “Structured Outputs is a feature that ensures your model responses conform to your provided JSON Schema”
   • [claimed-docs] “With `strict: true`, the model uses constrained decoding to guarantee that the output will always match your schema exactly”
   • [claimed-docs] “To use tools, the model must be provided with tool definitions. These tool definitions are in JSON schema format and are passed to the model…”

   proof ↗⚑ flag

   Together AI✓fullclaimed8/10

   Together AI's docs explicitly describe structured outputs where supported models return JSON conforming to a supplied schema, avoiding retries/fragile parsing, alongside function/tool calling and multi-step agentic tool call chaining. This directly matches the story's request for JSON-schema-enforced structured outputs; missing for 10: no independent/hands-on validation of schema-conformance reliability or grammar-based constraints beyond JSON schema.

   • [claimed-docs] “Supported models can return JSON that conforms to any schema you supply, so you can read the output directly in code without retries or frag…”
   • [claimed-docs] “Function calling (also called tool calling) lets LLMs respond with structured function names and arguments that you can execute in your appl…”
   • [claimed-docs] “To build agent loops, chain tool calls inside one response (multi-step), and conversations that thread tools across many turns (multi-turn).”
   • [claimed-docs] “Multi-step function calling chains sequential function calls within one conversation turn.”

   proof ↗⚑ flag

Tools

1. ai-native userRely on faithful function/tool calling — including parallel and multi-step tool use — so agent loops run on open models without breaking

   weight 3 · round to Together AI
   Groq~partialcommunity5/10

   Groq's docs clearly document OpenAI-compatible function/tool calling (JSON-schema tool definitions, strict-mode structured outputs, and now Remote MCP where 'Groq handles all orchestration') which supports agentic tool use, but none of the evidence explicitly confirms parallel tool calls or multi-step tool-use loops working reliably. Community reports also raise reliability concerns ('so many random errors, so many silly quirks') that create some doubt about robustness for agent loops, though not a concrete tool-calling failure. missing for 10: explicit documentation/examples of parallel tool calling, multi-step agentic loop reliability testing, and independent hands-on verification that tool calls remain faithful across steps.

   • [claimed-docs] “Tool use (or function calling) is what transforms a language model from a conversational interface into an autonomous agent capable of takin…”
   • [claimed-docs] “To use tools, the model must be provided with tool definitions. These tool definitions are in JSON schema format and are passed to the model…”
   • [claimed-docs] “With strict: true, the model uses constrained decoding to guarantee that the output will always match your schema exactly.”
   • [claimed-docs] “point to an MCP server URL and the Groq API will start using its tools without you having to implement any tool logic yourself”
   • [claimed-docs] “Groq's Responses API supports remote tool use via MCP servers via HTTPS where Groq handles all orchestration... You don't implement anything…”
   • [community] “My company had a really terrible experience trying to use Groq, and I would NOT recommend anyone use their service if you need reliability. …”
   • [community] “There's a trail of complaints going back years now, and they rounded out the bottom of Kimi's verification program. Groq hosted models were/…”

   proof ↗⚑ flag

   Together AI~partialclaimed6/10

   Together AI documents native function/tool calling with explicit support for multi-step (chained) and multi-turn agentic tool use, positioned as the foundation for agent loops (together-ai-docs-8,9,15,16). However, there is no independent/hands-on evidence confirming reliability of parallel tool calls or that it avoids breakage across open models in practice — only first-party docs. missing for 10: independent benchmarks or community reports on parallel tool-call accuracy, explicit parallel tool-calling documentation, and hands-on verification that agent loops don't break on open models.

   • [claimed-docs] “Function calling (also called tool calling) lets LLMs respond with structured function names and arguments that you can execute in your appl…”
   • [claimed-docs] “To build agent loops, chain tool calls inside one response (multi-step), and conversations that thread tools across many turns (multi-turn).”
   • [claimed-docs] “Function calling (also called *tool calling*) lets LLMs respond with structured function names and arguments that you can execute in your ap…”
   • [claimed-docs] “Multi-step function calling chains sequential function calls within one conversation turn.”

   proof ↗⚑ flag