Table of Contents

πŸ‡«πŸ‡· Version franΓ§aise

LLM Provider Comparison β€” Orkeon

Derived from the source code (src/core/Orkeon.Infrastructure/LLMs/) and from each provider's declared LlmProviderCapabilities; campaign figures carry their own date. Legend: βœ“ supported Β· βœ— absent Β· ◐ partial/generic. Every provider but Azure OpenAI has an archived real-execution campaign; OpenRouter and Mammouth AI since 2026-10-07.

Provider Base class SSE streaming Native tool calling Multi-turn chat (tool roles) System message top_p / stop GBNF grammar response_format thinking Vision reasoning_content round-trip Prompt cache Timing metrics Polly resilience API key sanitization
OpenAI OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ schema βœ“ effort βœ“ βœ— ◐ auto βœ— βœ“ βœ“
Azure OpenAI OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ schema βœ“ effort βœ“ βœ— ◐ auto βœ— βœ“ βœ“
Anthropic HttpLlmProviderBase βœ“ native βœ“ βœ“ βœ“ (native, separate) βœ“ βœ— βœ“ schema βœ“ toggle (adaptive, off per model) βœ“ βœ— βœ“ explicit βœ— βœ“ βœ“
DeepSeek OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ object βœ“ toggle βœ“ βœ“ βœ“ metrics βœ— βœ“ βœ“
Z.AI (GLM) OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ object βœ“ toggle βœ“ βœ— ◐ auto βœ— βœ“ βœ“
Together AI OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ schema βœ“ toggle (reasoning object) βœ“ βœ— ◐ auto βœ— βœ“ βœ“
Mistral AI OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ schema βœ“ effort βœ“ βœ— ◐ auto βœ— βœ“ βœ“
Qwen OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ object βœ“ budget βœ“ βœ— ◐ auto βœ— βœ“ βœ“
Kimi / Moonshot OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ object βœ“ toggle βœ“ βœ— ◐ auto βœ— βœ“ βœ“
Google Gemini OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ schema βœ“ effort βœ“ βœ— ◐ auto βœ— βœ“ βœ“
Grok (x.AI) OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ schema βœ“ effort βœ“ βœ— ◐ auto βœ— βœ“ βœ“
MiniMax OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ— (accepted but non-binding β€” measured) βœ“ toggle (off only; M3.1-Flash always reasons) βœ“ βœ“ ◐ auto βœ— βœ“ βœ“
HuggingFace OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ object βœ“ toggle (chat_template_kwargs) βœ“ βœ— ◐ auto βœ— βœ“ βœ“
OpenRouter OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ“ schema (per endpoint) βœ“ budget (reasoning object) βœ“ (per model) βœ— ◐ auto (+ cache_write_tokens) βœ— (usage.cost exposed) βœ“ βœ“
Mammouth AI OpenAI-compat βœ“ βœ“ βœ“ βœ“ βœ“ ◐ Llm:Grammar βœ— (accepted but non-binding β€” measured) βœ“ effort (reasoning_effort) βœ“ (per model) βœ— ◐ auto βœ— βœ“ βœ“
Ollama HttpLlmProviderBase βœ“ βœ“ (/api/chat) βœ“ (/api/chat) βœ“ (prepend) βœ“ (options) ◐ Llm:Grammar βœ“ schema βœ“ toggle βœ“ (images) βœ— βœ— ◐ (total_duration, eval_duration on /api/generate) βœ“ βœ“

How to read the capability columns

response_format, thinking and Vision are not hand-maintained here: each provider declares an LlmProviderCapabilities value object, and OpenAICompatibleProviderBase translates it into the OpenAI dialect once. Anthropic and Ollama write their own dialect (they do not extend the OpenAI-compatible base); Qwen and OpenRouter override the thinking translation for DashScope's fields and OpenRouter's reasoning object. The response format guide details the response_format surfaces.

  • response_format β€” object means the API guarantees well-formed JSON; schema means it validates against a JSON Schema server-side. A schema sent to an object-only provider is downgraded with a warning, never in silence. Anthropic is schema-only: it has no equivalent of json_object, so a schema-less JSON request there is reported rather than sent.
  • thinking β€” effort accepts a level hint only; toggle can also switch reasoning on and off; budget additionally accepts an explicit token budget (Qwen and OpenRouter only β€” Anthropic rejects budget_tokens with a 400 on the current generation). On the wire, Anthropic's "on" is written as thinking: {type: adaptive, display: summarized} and its "off" as the model's own value (between_tools on Sonnet 5.5, disabled elsewhere) β€” a payload detail, not a capability level.
  • Anything a provider does not support is reported. An option declared in YAML on a provider that cannot honour it produces an actionable warning naming the option, the provider and the remedy.

Notes

  • ◐ prompt cache (auto): the vendor caches prompt prefixes implicitly and reports the hit; OpenAICompatibleProviderBase.ParseSuccessResponse reads prompt_cache_hit/miss_tokens and the OpenAI-standard prompt_tokens_details.cached_tokens generically. βœ“ explicit (Anthropic) means the cache does nothing until a cache_control breakpoint is placed β€” opt in with LlmCacheConfig / the YAML cache: block of an agent or of a task (llm_override:). βœ“ metrics (DeepSeek) adds the vendor's own prompt_cache_hit_tokens / prompt_cache_miss_tokens to the response metadata.
  • Anthropic SSE: ChatStreamingAsync parses the Messages API event stream natively: tokens arrive as they are produced.
  • Ollama tool calling: goes through /api/chat when the conversation declares tools, replays tool calls, or carries a non-text part; everything else keeps /api/generate. An audio or file part is refused there with NotSupportedException, as on OpenAI and Anthropic. The text-fallback protocol remains in charge for models without tool support.
  • Azure OpenAI: two API shapes β€” the dated deployment URL (default) and the v1 GA surface (api_version: v1), which is the only path to the Responses API and to the non-OpenAI models Azure resells.
  • HuggingFace: model identifiers accept a routing suffix (:fastest / :cheapest / :preferred / :<partner>) β€” the only cost and latency lever on Inference Providers.
  • top_p / stop: written only when set, whatever the value β€” temperature too; nothing set, nothing written and the model applies its own default. Mistral writes top_p: 1 when nothing sets one. Ollama writes them in its options, with seed. No dialect writes frequency_penalty or presence_penalty, and none but Ollama writes seed: setting one produces the structured warning naming it (event id 110), never a silent drop.
  • GBNF grammar: no vendor API documents a grammar field, so no provider declares the capability (LlmProviderCapabilities.GbnfGrammar). ◐ Llm:Grammar means the field is written β€” on both of Ollama's endpoints, and by both payload builders of the OpenAI-compatible base β€” only when the settings say the endpoint behind Llm:BaseUrl takes it ("Llm": { "Grammar": true }, a llama.cpp-compatible server such as Docker Model Runner or llama-server). Otherwise a grammar is dropped with a structured warning naming that key. Anthropic's Messages API has no such field, whatever the key says. A structured_output deliverable sends its schema as a non-strict json_schema response format to a provider that declares βœ“ schema (see LLM response format).
  • OpenAI max_completion_tokens: OpenAI retired max_tokens on its current models (the 2026-08-30 campaign failed ten modes on that one field), so the OpenAI dialect writes max_completion_tokens β€” accepted by the older generations too (verified on gpt-4o-mini the same day). The compatible vendors keep max_tokens: the retirement is OpenAI's alone.
  • Gemini response_format: undocumented on the compat surface when first audited (2026-08-18) and undeclared then; measured live on 2026-08-30, the surface accepts json_object and json_schema and enforces the schema server-side; the provider declares JsonSchema.
  • DeepSeek vision: arrived with deepseek-v4-flash-vision-exp (measured 2026-08-30) and is native on the Flash tier since V4.1 Flash (2026-09-10): the default deepseek-flash sees, the experimental companion is retired. Per provider vs per model as everywhere: deepseek-v4-pro stays text-only and answers an image with the vendor's own error. deepseek-v4-flash is a retired model's name the API "temporarily" routes to V4.1 Flash β€” the default moved to the vendor's own id on 2026-09-19.
  • MiniMax: campaign-backed since 2026-08-30 (7/2/3, same day it was integrated). The reasoning arrives INLINE β€” every reply opens with a <think> block inside content, no separate field β€” and the dialect splits it out to reasoning_content, re-inlining it on replay as the vendor documents. response_format is accepted but NON-BINDING (a schema is ignored, json_object arrives fenced in markdown): the None declaration is a measurement (raw calls, 2026-08-30, recorded in the campaign catalog's MiniMax note β€” the archived M8 shows βž– because the None declaration keeps the option from ever being sent). Vision is per model: MiniMax-M2 answers "I'm unable to view the image"; MiniMax-M3, the default since 2026-11-11, reads the image (campaigns of 2026-09-21 and 2026-11-11).
  • Grok (x.AI): every declared capability is a live measurement β€” a full 12-mode campaign passed against api.x.ai through the generic OpenAI dialect before the provider class existed (2026-08-30, archived under llmproviders-test/custom-endpoints/). Keys carry the xai- prefix, which the factory infers.
  • OpenRouter: the model marketplace (465 models on 2026-10-07) behind one key, integrated documentation-first and campaigned on 2026-10-07 (11/1 on google/gemini-3.7-flash, the red being M10; the default has been google/gemini-3.8-flash since 2026-11-11, 12/0/0). Identifiers are vendor/model (mandatory prefix), with the :free / :nitro / :floor suffixes and the openrouter/auto router slug (usable, refused as a default: the served model drifts β€” the served_model metadata says who answered). The reasoning trace comes back as reasoning, never reasoning_content (a ReasoningFieldName hook on the base); thinking travels as the reasoning request object (enabled / effort / max_tokens, so the Budget declaration is one of the transport β€” OpenRouter converts effort and budget into each other per model); the real charge arrives in usage.cost (the cost metadata) with its upstream_inference_cost / is_byok / cache_write_tokens / reasoning_tokens breakdown; two constant attribution headers (HTTP-Referer, X-OpenRouter-Title) name Orkeon. json_schema is honoured per endpoint and the provider does not send provider.require_parameters: whether a schema can be silently ignored elsewhere is still open after the first campaign (M8 green on the three models played). The advanced routing body (provider {…}, models[], plugins[]) is not exposed. The sk-or-v1- key prefix is confirmed on a real key (2026-10-07); the factory does not infer from it yet. The campaign also confirmed usage.cost on both paths, message.reasoning with reasoning_details[] on a reasoning model, and openrouter/auto echoing the model that answered.
  • Mammouth AI: the French multi-model subscription whose included API credits drive Orkeon, integrated documentation-first and campaigned on 2026-10-07 (10/0/2 on gemini-3.7-flash: every exercised mode green; the default has been gemini-3.8-flash since 2026-11-11, 9/1/2 with M10 red). On three concordant clues (2026-09-18) the API is a LiteLLM proxy; nothing in the provider depends on it. Identifiers are the vendors' own bare strings (gpt-5.6-sol, claude-sonnet-5, gemini-3.8-flash), so the provider is reached by host (api.mammouth.ai) or by the explicit provider key mammouth (LlmProviderFactory.Create("mammouth", …), orkeon llm probe -p mammouth) and never inferred from a model name β€” the same string without a base URL keeps going to the vendor. A run's settings carry no provider key, so a run reaches Mammouth through its BaseUrl. Only messages, model, temperature, max_tokens, top_p and stream are documented. Measured on 2026-10-07: tools are passed through, response_format is accepted but a strict schema does not bind (the MiniMax case, so it stays undeclared), reasoning_effort is accepted and the trace of a reasoning model comes back as reasoning_content; the provider declares the effort (reasoning_effort, the five levels) since #356, and the capability table holds Mammouth's own values for Gemini 3.7 and 3.8 (minimal to max, none refused; campaign M18, 2026-11-13); vision is declared from the vendor's own text, image model list. Prices are the vendor's upper bounds (gemini-3.8-flash, like gemini-3.7-flash before it, at 1.5 / 7.5 $/M, twice the direct price).
  • Anthropic identity-linked keys: refuse every request without an anthropic-workspace-id header (2026-08-30). Set LlmConfig.WorkspaceId (CLI: --workspace-id); classic keys need nothing.
  • Polly and API key sanitization: provided by HttpLlmProviderBase β†’ active everywhere.

API keys: the variable per provider

A run takes the key the configuration resolves, else the variable the settings file names: Llm:ApiKey in the file, overridden by ORKEON_Llm__ApiKey in the environment (ORKEON_Llm__Profiles__<id>__ApiKey for a profile); otherwise the variable the section's ApiKeyEnvVar names β€” read in the process environment, then, on Windows, in the user scope (configuration). The vendor names below are the conventional ones: a settings file stores the name of the variable and never the key, as the ApiKeyEnvVar of the Llm section and of a profile's Llm:Profiles:<id> entry. So exporting DEEPSEEK_API_KEY is enough for a terminal orkeon run on a file that names it: "ApiKeyEnvVar": "DEEPSEEK_API_KEY". A file that names no variable still reads Llm:ApiKey and ORKEON_Llm__ApiKey only.

Provider Variable (the conventional name for ApiKeyEnvVar) Key issued at Suggested timeout Balance exposed by the API (details)
OpenAI OPENAI_API_KEY platform.openai.com/api-keys engine default (30 s) βœ— (spend only, to an admin key)
Anthropic ANTHROPIC_API_KEY console.anthropic.com engine default βœ— (spend only, to an Admin API key)
DeepSeek DEEPSEEK_API_KEY platform.deepseek.com 600 s βœ“ GET /user/balance
Mistral AI MISTRAL_API_KEY console.mistral.ai engine default βœ— (usage only, to an Admin API key)
Google Gemini GEMINI_API_KEY aistudio.google.com/apikey engine default βœ—
Grok (x.AI) XAI_API_KEY console.x.ai engine default ◐ to a management key only
MiniMax MINIMAX_API_KEY platform.minimax.io 600 s βœ—
Together AI TOGETHER_API_KEY api.together.ai engine default βœ—
Qwen DASHSCOPE_API_KEY dashscope.console.aliyun.com engine default ◐ to an Alibaba Cloud AccessKey only
Kimi (Moonshot) MOONSHOT_API_KEY platform.moonshot.ai 600 s βœ“ GET /v1/users/me/balance (USD on .ai, CNY on .cn)
HuggingFace HF_TOKEN huggingface.co/settings/tokens engine default βœ—
Z.AI (GLM) ZAI_API_KEY z.ai/manage-apikey 600 s βœ—
OpenRouter OPENROUTER_API_KEY openrouter.ai/keys engine default ◐ per key (GET /api/v1/key); the account's credits to a management key only
Mammouth AI MAMMOUTH_API_KEY mammouth.ai β€” the vendor documents "from the API settings"; the exact page is confirmed with the first key engine default βœ— (a key's spend, answer undocumented)
Ollama Β· Docker Model Runner none β€” engine default β€” (local, no account)
OpenAI-compatible (custom) none of its own: the settings name it through ApiKeyEnvVar, else ORKEON_Llm__ApiKey is read, the default key of every run β€” engine default βœ— (unknown host)
Azure OpenAI no dedicated variable: its per-resource endpoint makes it an OpenAI-compatible entry β€” β€” βœ— (spend in Cost Management, to a Microsoft Entra identity)

The suggested 600 s is not decorative. The four providers whose default model reasons before it answers overrun the engine's 30 s, and the run of 2026-09-20 was lost to two Kimi timeouts reported as an empty answer. A settings file needs Llm:TimeoutSeconds raised explicitly.

Two names that are not this one, and get confused with it:

  • ORKEON_LLM_API_KEY β€” the default of orkeon llm probe -k and orkeon llm models -k. Campaign tooling only; a run never reads it.
  • ORKEON_<NAME> β€” the secret chain of the tools, not of the LLM (EnvironmentSecretProvider, then Secrets:<NAME> in the file): ORKEON_TAVILY_API_KEY for web_search, BRAVE_API_KEY read as-is for brave_search. See the configuration reference.

orkeon init --api-key-env <name> writes Llm:ApiKeyEnvVar: <name> into the generated file β€” nothing for ORKEON_Llm__ApiKey, which the runtime reads natively β€” so every run reads the variable it names.

Account balance: what each API tells

What is left on the account behind a key, as each vendor's API tells it. Only three vendors serve a balance to the inference key a run holds; for every other provider, the vendor's console is the place to look β€” the "Key issued at" column above, behind https://. A balance request goes to the key's own host: a key of Kimi's .cn platform is refused by the .ai host, and the other way round. Each row was read on the vendor's own reference on 2026-09-24.

Provider Balance Endpoint What it returns Key it takes Source
DeepSeek βœ“ available GET /user/balance balance_infos[], one entry per currency (CNY, USD): total_balance, granted_balance, topped_up_balance as decimal strings; is_available the API key (Bearer) api-docs.deepseek.com/api/get-user-balance
Kimi (Moonshot) βœ“ available GET /v1/users/me/balance data.available_balance, voucher_balance, cash_balance (negative in arrears) β€” US dollars on api.moonshot.ai, yuan on api.moonshot.cn; the answer names no currency the API key (Bearer) of that platform platform.kimi.ai/docs/api/balance Β· platform.kimi.com/docs/api/balance
OpenRouter ◐ per key GET /api/v1/key data.limit_remaining: US dollars the key may still spend under its limit, null when it has none; limit, usage. The account's credits (total_credits, total_usage) are on GET /api/v1/credits the API key for /key; a management key for /credits, which answers 403 to any other openrouter.ai/docs/api/api-reference/api-keys/get-current-api-key Β· openrouter.ai/docs/api/api-reference/credits/get-remaining-credits
Grok (x.AI) ◐ admin key required GET /v1/billing/teams/{team_id}/prepaid/balance on management-api.x.ai the prepaid balance, in US cents a management key only docs.x.ai/developers/rest-api-reference/management/billing
Qwen ◐ admin key required QueryAccountBalance, Alibaba Cloud billing OpenAPI AvailableAmount, AvailableCashAmount, CreditAmount, Currency an Alibaba Cloud AccessKey allowed bss:DescribeAcccount β€” never the DashScope key alibabacloud.com/help/en/user-center/developer-reference/api-bssopenapi-2017-12-14-queryaccountbalance
OpenAI βœ— not exposed none β€” GET /v1/organization/costs reports spend spend per day (amount.value, currency), not a balance an admin key developers.openai.com/cookbook/examples/completions_usage_api
Anthropic βœ— not exposed none β€” GET /v1/organizations/cost_report reports spend spend in US cents, not a balance Admin API credentials β€” "workspace API keys don't work" platform.claude.com/docs/en/manage-claude/usage-cost-api
Mistral AI βœ— not exposed none β€” GET /v1/admin/usage reports usage consumption per category for a month; the credit balance is on the console's Billing page only an Admin API key docs.mistral.ai/admin/admin-api/usage-metrics Β· docs.mistral.ai/admin/billing-usage/billing
Azure OpenAI βœ— not exposed none β€” Cost Management reports spend spend per meter a Microsoft Entra identity with Cost Management Reader, never the resource key learn.microsoft.com/en-us/azure/foundry/concepts/manage-costs
Google Gemini βœ— not exposed none the prepay balance is managed "directly within the Google AI Studio Billing tab" β€” ai.google.dev/gemini-api/docs/billing
Together AI βœ— not exposed none the "Credit balance" card of the billing settings β€” docs.together.ai/docs/billing-credits Β· docs.together.ai/docs/billing-usage-limits
HuggingFace βœ— not exposed none the billing page and the Inference Providers settings β€” huggingface.co/docs/inference-providers/pricing
Z.AI (GLM) βœ— not exposed none the API reference lists no billing endpoint β€” docs.z.ai/llms.txt
MiniMax βœ— not exposed none for pay-as-you-go; GET /v1/token_plan/remains on www.minimax.io for a Token Plan Token Plan only, answer undocumented a Token Plan subscription key platform.minimax.io/docs/token-plan/faq
Mammouth AI βœ— not exposed none β€” GET /key/info, documented with a placeholder host "how much credits has been spent on a key": spend, answer undocumented the API key info.mammouth.ai/docs/api-quick-start
Ollama Β· Docker Model Runner β€” not applicable none a local runtime has no account β€” β€”

Two endpoints are documented without the shape of their answer: MiniMax's Token Plan remains and Mammouth's key/info. An OpenAI-compatible endpoint (custom) exposes no balance anyone knows of, and a local one has no account.

Defaults and newer models β€” catalogue reviews of 2026-09-19, 2026-10-07 and 2026-11-11

Every vendor's own model and pricing pages were read on 2026-09-19, alongside the public catalogues of the two aggregators (OpenRouter, 447 models; Mammouth, 100). The rule is the one LlmProviderDefaultModels states: a default changes only when the vendor retires the name; otherwise a newer model is a candidate until a campaign has archived an M1 on it. The campaigns of 2026-10-07 and 2026-11-11 played the candidates again: the date in a verdict says which campaign measured what, and the models they added to the candidates are in the "Newer" column.

Provider Default (code) Served on 2026-09-19 Newer on the vendor's API Verdict
OpenAI gpt-5.6-sol yes β€” 4 / 20 $/M through 2026-11-21, the replacement target of both 2026 deprecation waves gpt-6-astra (10 / 50, effort low…max with no none, 2Γ— billing above 272K input tokens); gpt-6.1-terra (listed after 2026-10-07) stays β€” campaigned 2026-09-21: gpt-6-astra 10/2 with temperature: 1 pinned; it has no reasoning_effort: none, so function tools on /v1/chat/completions are refused (use /v1/responses) β€” not a default until the dialect speaks Responses; gpt-6-astra-pro is an aggregator label, not an OpenAI id; campaigned 2026-11-11 (after the fix that lets reasoning_effort: none reach the API again): gpt-5.6-sol 11/1, gpt-6-sol 11/1, gpt-6-luna 12/0/0, gpt-6.1-terra 12/0/0 β€” it takes none, unlike gpt-6.1-sol; gpt-6-astra 10/2 and gpt-6.1-sol 9/3 stay red on M2 and M5 (no none, tools need /v1/responses)
Anthropic claude-sonnet-5 yes β€” 2 / 10 $/M made permanent, active until at least 2027-06-30 claude-fable-5-1 (2026-09-01, 10 / 50, thinking always on, forced tool_choice refused); claude-fable-5-2 (2026-11-09, 1M context, 128K output) stays β€” campaigned 2026-09-21: claude-fable-5-1 11/1 twice (the red is M6, the text-fallback protocol; native tools, schema, vision and cache all green) β€” viable, at five times the price; claude-opus-5-fast is a speed flag, not an id; temperature / top_p / top_k return 400 on every model from Opus 4.7 on; campaigned 2026-11-11: the default 12/12, claude-fable-5-2 12/12 β€” including the M6 that claude-fable-5-1 still misses (11/1)
Azure OpenAI deployment β€” β€” nothing to default to
Ollama llama3.2 yes β€” 3B, text only, no thinking, a year old qwen3.5:4b, gemma4:e4b, granite4.2:3b (tools + thinking, same size class) stays: a new default means a pull on every machine
Together AI zai-org/GLM-5.3-Flash (was meta-llama/Llama-3.3-70B-Instruct-Turbo) yes β€” 0.15 / 0.50 $/M (catalogue read of 2026-11-11); the former default is no longer served on serverless Qwen/Qwen3.5-9B, deepseek-ai/DeepSeek-V4.1-Flash; Llama 4 left serverless changed 2026-11-11 β€” every call to the former default answered Unable to access non-serverless model (2/9/1), so the default moves to zai-org/GLM-5.3-Flash, Together's replacement target since 2026-09-19: 11/0/1 on the last run of the day (10/1/1 on the first, M10: no cached tokens reported; 11/0/1 on 2026-09-21), it reads the image, so the default sees and there is no separate vision companion; Qwen/Qwen3.5-9B 10/1/1 (M10), candidate
DeepSeek deepseek-flash (was deepseek-v4-flash) yes β€” V4.1 Flash, 0.30 / 1.20 $/M peak, native vision it is the newest changed 2026-09-19: the old name is a retired model "temporarily" routed here; replayed 2026-09-21: 12/12 twice β€” M2 and M9, red for five campaigns on deepseek-v4-flash, are green on deepseek-flash; replayed 2026-11-11: deepseek-flash 12/12
Kimi kimi-k3.1-pro (was kimi-k2.6) yes β€” kimi-k2.6 0.95 / 4.00 $/M, no retirement date kimi-k3 (July 2026, 1M, 3 / 15, fixed sampling; thinking: disabled accepted, campaign M18 2026-11-13); kimi-k3.1-pro and the -highspeed variants (listed 2026-11-11) changed 2026-11-11 β€” campaigned 2026-09-21: kimi-k3 12/12 then 11/1 (M2 on the flattened shape), reasoning trace, cache 0.97, bare JSON where k2.6 now fences its json_object β€” ready, at three times the price; kimi-k2.5 and moonshot-v1-* retired 2026-08-31; docs now on platform.kimi.ai; 2026-11-11: kimi-k2.6 10/1/1 (M8: its json_object still fenced), kimi-k3 12/12, kimi-k3.1-pro 12/12 β€” the new top model; changed 2026-11-11: the default moves to kimi-k3.1-pro (12/12 direct and on OpenRouter). The K3 family fixes temperature, top_p and its reasoning: KimiLlmProvider leaves temperature and top_p out of a K3 request and reports each, so a host that sets them gets a warning, not a 400; the thinking switch is sent, since K3 honours it (M18, 2026-11-13)
Qwen qwen3.7-plus yes β€” still the Plus tier, one of the three recommended models qwen3.8-max (= qwen3.8-max-0902), qwen3.8-flash; no qwen3.8-plus stays β€” campaigned 2026-09-21: qwen3.8-max 12/12, qwen3.8-flash 11/1 (M2), both ready, and qwen3.7-plus is still 12/12; the 3.8 generation adds preserve_thinking; 2026-11-11: qwen3.7-plus, qwen3.8-flash and qwen3.8-max 12/12 each
Mistral AI mistral-medium-2604 yes β€” 1.50 / 7.50 $/M, not deprecated nothing for chat since April (OCR 4.1 only) stays; devstral-* / magistral ids sit in the deprecated table; 2026-11-11: the default and mistral-large-4 12/12, mistral-small-2603 11/1 (M9)
HuggingFace meta-llama/Llama-3.1-8B-Instruct yes β€” but tools on one of its four routed providers, and :fastest may pick another Qwen/Qwen3.5-9B (tools on three providers, from 0.10 / 0.15), zai-org/GLM-5.3-Flash, deepseek-ai/DeepSeek-V4.1-Flash candidate Qwen3.5-9B β€” campaigned 2026-09-21: 10/0/2, it sees and its tools hold; its first pass returned an empty M1 after 4113 tokens of thinking on the router's 4096 fallback β€” pin Llm:MaxTokens (or turn thinking off) before making it a default. The routing still explains the default's moving reds (M2 and M5 both red on the same pass); 2026-11-11: the default 7/3/2 (M2, M5, M9), Qwen/Qwen3.5-9B 9/1/2 (M2)
Z.AI glm-5.3-flashx (was glm-5.2) yes β€” glm-5.2 still under "Latest Models", 1.40 / 4.40 $/M glm-5.3 (2026-08-18), glm-5.3-flash (0.15 / 0.50, image + video), glm-5.3-flashx changed 2026-11-11 β€” campaigned 2026-09-21: glm-5.3-flash 12/12 twice (it sees, where 5.2 and 5.3 answer 1210; cache read) β€” the cleanest candidate in the fleet, at a ninth of the price; glm-5.3 10/2 (M2, M9) brings nothing over 5.2. The Toggle declaration still needs a per-model guard before a bump; 2026-11-11: glm-5.3-flashx 12/12 again, glm-5.3 11/1 (M9), glm-5.3-flash 11/1 (M2), glm-5.2 10/2; changed 2026-11-11: the default moves to glm-5.3-flashx (12/12 on 2026-10-07 and on the first run of 2026-11-11, 11/1 with M10 red on the last; it sees and its cache is read); last runs of the day: glm-5.3-flash 12/0/0, glm-5.3 10/2 (M2, M9), glm-5.2 11/1 (M9). The 5.3 family cannot switch thinking off: ZaiLlmProvider does not send thinking: disabled to it and reports the option instead
Google Gemini gemini-3.8-flash (was gemini-3.7-flash) yes β€” gemini-3.7-flash "previous generation", no shutdown date, 0.75 / 3.75 $/M until 2026-12-31 then 1.50 / 7.50 gemini-3.8-flash (GA 2026-09-02, same price, minimal thinking rejected); gemini-4-argon, gemini-4-carbon first candidate for a bump β€” campaigned 2026-09-21: gemini-3.8-flash 11/0/1 twice, identical to 3.7 mode for mode on the direct transport; OpenRouter and Mammouth not played (no key); 2026-11-11: the default 11/0/1, gemini-3.8-flash 12/0/0, gemini-4-argon and gemini-4-carbon 11/0/1 each; gemini-omni-1.2-flash is listed but Interactions-API only, so out of reach of this provider and not campaigned; changed 2026-11-11 (full rerun): gemini-3.8-flash 12/0/0 direct and on OpenRouter (9/1/2 on Mammouth, the same M10 as 3.7), the default 11/0/1 β€” the default moves to gemini-3.8-flash, and OpenRouter and Mammouth follow it
Grok (x.AI) grok-4.6 yes β€” recommended, 2 / 6 $/M below 200K prompt tokens, 4 / 12 above none stays (grok-4.7 is listed since); campaigned 2026-11-11: grok-4.6 and grok-4.7 12/12
MiniMax MiniMax-M3 (was MiniMax-M2) yes β€” MiniMax-M2 listed as legacy, no retirement date MiniMax-M3 (2026-06-01, 1M, image + video input, same 0.30 / 1.20); MiniMax-M3.1-Flash changed 2026-11-11 β€” campaigned 2026-09-21: MiniMax-M3 9/1/2 then 8/1/3 β€” it reads the image (then the kit's vision companion), M2 red like M2, the inline <think> block split out as on M2, 128 cached tokens once; the reasoning-format question is answered; 2026-11-11: MiniMax-M3.1-Flash 10/0/2 (M7 and M8 not declared by the provider), M9 green β€” it reads the image where MiniMax-M3 reads 72; MiniMax-M3 9/1/2 (M9), MiniMax-M2 8/2/2 (M2, M9); changed 2026-11-11 (full rerun): MiniMax-M3 10/0/2 (it reads the image), MiniMax-M2 6/4/2 (M2, M5, M9, M10) β€” the default moves to MiniMax-M3 and the vision companion is retired
OpenRouter google/gemini-3.8-flash (was google/gemini-3.7-flash) yes google/gemini-3.8-flash (2026-09-02, same price) follows the direct default β€” campaigned 2026-10-07: 11/1 on both (M10: cache fields returned with 0); 2026-11-11: the default 12/0/0, google/gemini-3.8-flash 11/1/0 (M10), moonshotai/kimi-k3.1-pro 12/0/0; changed 2026-11-11 (full rerun): google/gemini-3.8-flash 12/0/0, google/gemini-3.7-flash 11/1 (M10) β€” follows the direct default to 3.8
Mammouth AI gemini-3.8-flash (was gemini-3.7-flash) yes gemini-3.8-flash follows the direct default β€” campaigned 2026-10-07: 10/0/2 on both, every exercised mode green; 2026-11-11: 9/1/2 on both (M10: 0 cached tokens), kimi-k3.1-pro 10/0/2; changed 2026-11-11 (full rerun): 9/1/2 on both again (M10) β€” follows the direct default to 3.8

ModelPricingRegistry follows the same pages: the GPT-5.6 family and gpt-6-astra, and, for Anthropic (read 2026-10-30), Fable 5.1 / Fable 5 at 10 / 50, Opus 5.5 at 4 / 20, Opus 5 and Opus 4.5 to 4.8 at 5 / 25, Sonnet 5 at 2 / 10, Haiku 5 at 0.10 / 0.50 and Haiku 4.5 at 1 / 5. Cache tariffs: a 5-minute write at 1.25x the prompt rate, a 1-hour write at 2x (costed by the cost manager), a read at 0.1x (0.25 for Fable 5.1, 0.20 for Opus 5.5). The Haiku 5.5 tier above 100K prompt tokens (0.50 / 2.50) is not modelled.

Every paid provider's default model is priced (read 2026-11-15), $ per million input / output tokens: DeepSeek deepseek-flash 0.30 / 1.20 (the peak rate of this page; cache read 0.006); Z.AI glm-5.3-flashx 0.37 / 1.25, glm-5.3-flash 0.15 / 0.50 and glm-5.3 1.4 / 4.4 (docs.z.ai pricing); Kimi kimi-k3.1-pro 2 / 10 and kimi-k3 3 / 15 and Qwen qwen3.7-plus 0.4 / 1.6 (Alibaba Cloud Model Studio, Singapore: Kimi's own page shows no figure, and Qwen's rate is the limited-time one); Gemini gemini-3.8-flash 1.50 / 7.50 with a cache read at 0.15 (the rate Google applies from 2027-01-01, kept as the conservative figure for budgets) and gemini-3.7-flash 0.75 / 3.75 (no cache-read price); Mistral mistral-medium-2604 1.50 / 7.50; Grok grok-4.6 and grok-4.7 2 / 6; MiniMax MiniMax-M3 0.30 / 1.20. The gateways are priced by their own rows, scoped by provider: Together zai-org/GLM-5.3-Flash 0.15 / 0.50, Mammouth gemini-3.8-flash and gemini-3.7-flash 1.5 / 7.5 (the gateway's upper bound, from secondary sources). Ollama and Docker Model Runner models cost 0. Hugging Face stays unpriced: its routed provider sets the price, so a call warns once per model and names Orkeon:CostTracking:CustomPricings, where a row prices it. OpenRouter is counted from the vendor's own usage.cost, not from this list; an Azure OpenAI deployment is priced only when its name matches a row (add one under CustomPricings). Not modelled: the Qwen rate above 256K input tokens (1.2 / 4.8) and the Grok rate above 200K prompt tokens (4 / 12).

Models served and their default options β€” catalogue read of 2026-11-11

The latest result of every campaigned model, regenerated after each campaign, is on LLM campaign results; the scores below are the ones this catalogue read recorded.

The campaign of 2026-11-11 read each provider's own catalogue endpoint with the campaign key (orkeon llm models -p <provider> returns the same identifiers), then played the protocol on every default, on the candidates of the reviews, and on the current chat models the listing showed that no campaign had met yet. Two providers were campaigned later the same day: Kimi with a second key (the first was refused, 401), and Z.AI on a working key (the first account had no balance: 429, Insufficient balance or no resource package). A score marked (2026-10-07) below was not replayed. Three kinds of facts come out of it, kept apart on purpose:

  • what the endpoint publishes about a model β€” context, output cap, default sampling, effort levels. The vendor's own data: read, not measured;
  • what the model dictates β€” a value it refuses to see changed. No endpoint publishes that. It is learnt by calling the model (one pass on M1, M5 and M7 at the probe's defaults) and recorded in requiredParams;
  • what the campaign measured on the twelve automated modes.

The full listing β€” 1 312 identifiers across fourteen providers, each with the options its endpoint publishes β€” is llmproviders-test/lib/served-models.json, a dated snapshot the next campaign diffs against. Azure OpenAI and Ollama have no row: one serves an account's deployments, the other what a machine pulled.

Provider Catalogue request Identifiers What it publishes per model
OpenAI GET api.openai.com/v1/models 96 ids only
Anthropic GET api.anthropic.com/v1/models 15 context, output cap, effort levels, thinking types, modalities
DeepSeek GET api.deepseek.com/models 2 context, output cap, modalities, effort levels and default
Z.AI GET api.z.ai/api/paas/v4/models 12 ids only
Kimi GET api.moonshot.ai/v1/models 6 context, image/video input, reasoning
Google Gemini GET generativelanguage.googleapis.com/v1beta/models 62 context, output cap, default temperature / topP / topK, maximum temperature, thinking
Mistral AI GET api.mistral.ai/v1/models 64 context, default temperature, capabilities, aliases, deprecation
Qwen GET dashscope-intl.aliyuncs.com/compatible-mode/v1/models 101 ids only
Grok (x.AI) GET api.x.ai/v1/language-models 8 modalities, prices in $ per million tokens, long-context threshold, aliases
MiniMax GET api.minimax.io/v1/models 9 ids only
Together AI GET api.together.xyz/v1/models 270 type, context, prices in $ per million tokens
HuggingFace GET router.huggingface.co/v1/models 127 modalities, and per routed provider: context, tools, structured output
OpenRouter GET openrouter.ai/api/v1/models 433 context, output cap of the top endpoint, default parameters, reasoning (mandatory, default effort, supported efforts)
Mammouth AI GET api.mammouth.ai/v1/models 107 context, output cap

The chat models of the current generation, provider by provider. An em dash in the two size columns means the endpoint does not publish the figure β€” for those models the documented caps are the ones of the next section.

Provider Model Context Max output Published by the endpoint Dictated by the model (measured) Campaign βœ…/❌/βž–
OpenAI gpt-5.6-sol (default) β€” β€” β€” temperature = 1 Β· reasoning_effort = none with function tools 11/1/0 β€” ❌ M9
gpt-5.6-terra β€” β€” β€” temperature = 1 Β· reasoning_effort = none with function tools 11/1/0 β€” ❌ M9 (2026-10-07)
gpt-5.6-luna β€” β€” β€” temperature = 1 Β· reasoning_effort = none with function tools 11/1/0 β€” ❌ M6 (2026-10-07)
gpt-6-sol β€” β€” β€” temperature = 1 Β· reasoning_effort = none with function tools 11/1/0 β€” ❌ M9
gpt-6-luna β€” β€” β€” temperature = 1 Β· reasoning_effort = none with function tools 12/0/0
gpt-6.1-terra β€” β€” β€” temperature = 1 Β· reasoning_effort = none with function tools 12/0/0
gpt-6.1-sol β€” β€” β€” temperature = 1 9/3/0 β€” ❌ M2, M5, M9
gpt-6-astra β€” β€” β€” temperature = 1 10/2/0 β€” ❌ M2, M5
Anthropic claude-sonnet-5 (default) 1 000 000 128 000 efforts low medium high xhigh max Β· thinking adaptive disabled temperature = 1 12/0/0
claude-sonnet-5-5 1 000 000 128 000 efforts low medium high xhigh max Β· thinking adaptive temperature = 1 12/0/0
claude-haiku-5-5 1 000 000 128 000 efforts low medium high xhigh max Β· thinking adaptive disabled temperature = 1 12/0/0
claude-opus-5 1 000 000 128 000 efforts low medium high xhigh max Β· thinking adaptive disabled temperature = 1 12/0/0 (2026-10-07)
claude-opus-5-5 1 000 000 128 000 efforts low medium high xhigh max Β· thinking adaptive temperature = 1 12/0/0
claude-fable-5-1 1 000 000 128 000 efforts low medium high xhigh max Β· thinking adaptive temperature = 1 11/1/0 β€” ❌ M6
claude-fable-5-2 1 000 000 128 000 efforts low medium high xhigh max Β· thinking adaptive temperature = 1 12/0/0
DeepSeek deepseek-flash (default) 1 048 576 393 216 default effort high (of low high max) Β· text + image β€” 12/0/0
deepseek-v4-pro 1 048 576 393 216 default effort high (of low high max) Β· text β€” 11/1/0 β€” ❌ M9 (2026-10-07)
Z.AI glm-5.2 β€” β€” β€” β€” 11/1/0 β€” ❌ M9 (10/2/0 earlier the same day)
glm-5.3 β€” β€” β€” β€” 10/2/0 β€” ❌ M2, M9 (11/1/0 earlier the same day)
glm-5.3-flash β€” β€” β€” β€” 12/0/0 (11/1/0, M2 red, earlier the same day)
glm-5.3-flashx (default) β€” β€” β€” β€” 11/1/0 β€” ❌ M10 (12/0/0 earlier the same day)
Kimi kimi-k2.6 262 144 β€” image + video + reasoning temperature = 1 10/1/1 β€” ❌ M8
kimi-k3 1 048 576 β€” image + video + reasoning temperature = 1 11/1/0 (M2)
kimi-k3.1-pro (default) 1 048 576 β€” image + video + reasoning temperature = 1 12/0/0
kimi-k3.1-pro-highspeed 1 048 576 β€” image + video + reasoning temperature = 1 12/0/0
kimi-k2.7-code 262 144 β€” image + video + reasoning temperature = 1 12/0/0
kimi-k2.7-code-highspeed 262 144 β€” image + video + reasoning temperature = 1 12/0/0
Google Gemini gemini-3.7-flash 1 048 576 65 536 temperature 1 (max 2) Β· topP 0.95 Β· topK 64 β€” 11/0/1
gemini-3.8-flash (default) 1 048 576 65 536 temperature 1 (max 2) Β· topP 0.95 Β· topK 64 β€” 12/0/0
gemini-4-argon 1 048 576 32 768 temperature 1 (max 2) Β· topP 0.95 Β· topK 64 β€” 11/0/1
gemini-4-carbon 1 048 576 32 768 temperature 1 (max 2) Β· topP 0.95 Β· topK 64 β€” 11/0/1
gemini-3.5-flash-lite 1 048 576 65 536 temperature 1 (max 2) Β· topP 0.95 Β· topK 64 β€” 12/0/0 (2026-10-07)
gemini-3.1-pro-preview 1 048 576 65 536 temperature 1 (max 2) Β· topP 0.95 Β· topK 64 β€” 10/1/1 β€” ❌ M5 (2026-10-07)
Mistral AI mistral-medium-2604 (default) 262 144 β€” temperature 1 Β· tools + reasoning + vision reasoning_effort ∈ {high, none} 12/0/0
mistral-large-4 1 048 576 β€” temperature 1 Β· tools + reasoning + vision reasoning_effort ∈ {low, medium, high} 12/0/0
mistral-small-2603 262 144 β€” temperature 0.3 Β· tools + reasoning + vision reasoning_effort ∈ {high, none} 11/1/0 β€” ❌ M9
Qwen qwen3.7-plus (default) β€” β€” β€” β€” 12/0/0
qwen3.7-flash β€” β€” β€” β€” 11/1/0 β€” ❌ M2 (2026-10-07)
qwen3.7-max β€” β€” β€” β€” 11/1/0 β€” ❌ M9 (2026-10-07)
qwen3.8-flash β€” β€” β€” β€” 12/0/0
qwen3.8-max β€” β€” β€” β€” 12/0/0
deepseek-v4.1-flash-1017 β€” β€” β€” β€” 12/0/0
Grok (x.AI) grok-4.6 (default) β€” β€” text + image Β· 2 / 0.5 / 6 $/M in / cached / out β€” 12/0/0
grok-4.7 β€” β€” text + image Β· 2 / 0.5 / 6 $/M in / cached / out β€” 12/0/0
MiniMax MiniMax-M2 β€” β€” β€” β€” 6/4/2 β€” ❌ M2, M5, M9, M10
MiniMax-M2.7 β€” β€” β€” β€” 8/2/2 β€” ❌ M2, M9 (2026-10-07)
MiniMax-M3 (default) β€” β€” β€” β€” 10/0/2
MiniMax-M3.1-Flash β€” β€” β€” β€” 10/0/2
Together AI zai-org/GLM-5.3-Flash (default) 1 048 575 β€” 0.15 / 0.5 $/M β€” 11/0/1 (10/1/1, M10 red, earlier the same day)
Qwen/Qwen3.5-9B 262 144 β€” 0.17 / 0.25 $/M β€” 10/1/1 β€” ❌ M10
meta-llama/Llama-3.3-70B-Instruct-Turbo (former default) 131 072 β€” 1.04 / 1.04 $/M β€” 2/9/1 β€” no longer serverless
HuggingFace meta-llama/Llama-3.1-8B-Instruct (default) up to 131 072 β€” 3 routed provider(s), 1 with tools β€” 7/3/2 β€” ❌ M2, M5, M9
Qwen/Qwen3.5-9B up to 262 144 β€” 3 routed provider(s), 2 with tools β€” 9/1/2 β€” ❌ M2
OpenRouter google/gemini-3.7-flash 1 048 576 65 536 default effort medium Β· reasoning mandatory β€” 11/1/0 β€” ❌ M10
google/gemini-3.8-flash (default) 1 048 576 65 536 default effort medium Β· reasoning mandatory β€” 12/0/0
moonshotai/kimi-k3.1-pro 1 048 576 943 718 default effort max β€” 12/0/0
deepseek/deepseek-v4.1-flash 1 048 576 943 718 default effort high β€” 4/0/0 (M5, M7, M8, M10) (2026-10-07)
Mammouth AI gemini-3.7-flash 1 048 576 65 536 β€” β€” 9/1/2 β€” ❌ M10
gemini-3.8-flash (default) 1 048 576 65 536 β€” β€” 9/1/2 β€” ❌ M10
kimi-k3.1-pro 1 048 576 943 718 β€” β€” 10/0/2
deepseek-v4.1-flash 1 048 576 1 048 576 β€” β€” 1/0/2 (M5, M7, M8) (2026-10-07)

What the table does not say by itself:

  • OpenAI β€” the 6 generation is listed (gpt-6-sol, gpt-6-luna, gpt-6.1-sol, gpt-6.1-terra, beside gpt-6-astra). gpt-6-sol, gpt-6-luna, gpt-6.1-terra and the two 5.6 tiers not replayed behave like the default: temperature 1 only, and function tools on /v1/chat/completions need reasoning_effort: none. gpt-6.1-terra takes none, unlike gpt-6.1-sol, and is 12/12 on its first run. gpt-6.1-sol behaves like gpt-6-astra: it has no none, so function tools stay refused there (M2 on the messages array, M5) until the dialect speaks /v1/responses. The first pass of 2026-11-11 (06:10) was red on M2 and M5 for every OpenAI model: Orkeon had been lowering the pinned none to low since #242, and the gpt-5 row of the capability table caught the gpt-5.6-* ids. Fixed and replayed at 06:25: gpt-5.6-sol 11/1, gpt-6-sol 11/1, gpt-6-luna and gpt-6.1-terra 12/12; gpt-6-astra and gpt-6.1-sol stay red, as expected. M9 is the mode that moves at the mandated temperature: 29, 26 and 18 were read for 73 on 2026-10-07, 23, 43 and 23 on 2026-11-11 (gpt-5.6-sol, gpt-6-sol, gpt-6.1-sol).
  • Anthropic β€” claude-sonnet-5-5, claude-opus-5-5 and claude-haiku-5-5 are listed, the last on the day of the campaign. The four models campaigned for the first time refuse a pinned temperature in the words claude-sonnet-5 uses. claude-opus-5-5 answers the probe's tool prompt with HTTP 200, an empty content and stop_reason: "refusal" (category cyber): the provider does not read stop_reason and returns an empty answer β€” its M5 and M6 reds are that refusal, not tool calling. claude-haiku-5-5's M9 red is the probe's: the model answers Seventy-three red, the check looks for the digits. On 2026-11-11 claude-haiku-5-5 is 12/12 and claude-opus-5-5 is 12/12 with the same prompt: the classifier no longer meets the probe's wording. claude-fable-5-2 (listed 2026-11-09) refuses an explicit temperature like the rest of the 5 family and is 12/12 on its first run, including the M6 that claude-fable-5-1 still misses (11/1).
  • Mistral AI β€” the account is no longer rate-limited, so the default has a result again (11/1, its 2026-08-30 line). mistral-large-4 reasons by default and streams delta.content as an array of typed chunks: the text stream threw on it (M3) and the chat stream dropped the trace and the first words of the answer. Both streams read the two shapes since 2026-10-07; replayed on that fix the model was 11/1, M3 green, and it is 12/12 on 2026-11-11 (its published context went from 524 288 to 1 048 576). The default is 12/12 too. /v1/models now also lists models of other vendors served by Mistral β€” deepseek-v4-1-flash-1017 and zai-glm-5-3-flash, both 11/1 on their first run: the system message is lost on the messages-array shape of M2 while the flattened shape honours it.
  • Kimi β€” played an hour after the others: the account was suspended for insufficient balance at first (HTTP 429, exceeded_current_quota_error, which the retry policy treats as a rate limit). kimi-k2.6 still fences its json_object in markdown (M8, as on 2026-09-21); kimi-k2.7-code is the one Kimi model with every mode green. It mandates temperature 1 too, which only its streams showed: on a buffered call the provider retries once with the mandated value, the two streaming paths do not. 2026-11-11, with a second key (the first was refused, 401 Incorrect API key provided): the listing grows to six identifiers β€” kimi-k3.1-pro and the -highspeed variants of it and of kimi-k2.7-code. All three mandate temperature 1 as well (6/6 at 0: M2, M3, M4, M5, M9 and M10 refused), and are 12/12 once pinned; kimi-k3 is 12/12 too, kimi-k2.7-code still 12/12. kimi-k2.6 stays 10/1/1 β€” M8 still fenced, M10 reports no cache-token breakdown. kimi-k3.1-pro is 12/12 through OpenRouter as well, 10/0/2 through Mammouth.
  • Z.AI β€” glm-5.3-flashx is the second model of the family to pass every mode (on its first run of the day). The listing omits glm-4.6v-flash, which is served (the vision companion, M9 green the same day): a catalogue is not the whole offer. glm-5.3-flashx is 12/12 again on the first run, 11/1 on the last (M10 red); glm-5.3 turns M2 green on the first run (11/1, M9 1210 left) and red again on the last (10/2); glm-5.3-flash keeps M2 red on the first run, its system message ignored on both shapes β€” the model, not Orkeon β€” and is 12/12 on the last; glm-5.2 is 10/2 then 11/1 (M9 1210; M10 0 cached on the first run). The first key of the day had no balance (429, code 1113), which the provider retried as a transient limit for 21 minutes (issue #328). glm-5.3-flashx-legal is new in the listing, not campaigned.
  • Qwen β€” an international key is served by dashscope-intl.aliyuncs.com only; the compiled endpoint (dashscope.aliyuncs.com) answers it 401. Set Llm:BaseUrl. The international listing fell from 172 to 101 ids on 2026-11-11 (the qwen3 open-weight line and the qwen-vl / omni / tts families left it); qwen3.7-plus, qwen3.8-flash and qwen3.8-max are 12/12, and deepseek-v4.1-flash-1017, served there, is 12/12 on its first run.
  • Gemini β€” gemini-3.5-flash-lite is the one Gemini model whose cache breakdown came back (4077 cached tokens); the M5 red of gemini-3.1-pro-preview did not reproduce in three replays (not replayed on 2026-11-11). gemini-3.8-flash is 12/12 on 2026-11-11 and the default 11/0/1. gemini-4-argon and gemini-4-carbon (1 048 576 context, 32 768 output, thinking) are 11/0/1 each on their first run. gemini-omni-1.2-flash is listed by /v1beta/models but answers This model only supports Interactions API on the OpenAI-compatible surface: out of this provider's reach, not campaigned.
  • MiniMax β€” MiniMax-M3.1-Flash is new in the listing (9 ids): 10/0/2 on its first run, M7 and M8 not declared by the provider, M9 green β€” it reads the image where MiniMax-M3 misreads the number (72; 9/1/2 on the first run, 10/0/2 on the last: it reads the image there too). MiniMax-M2 is 8/2/2 (M2, M9), then 6/4/2 on the last run (M2, M5, M9, M10).
  • Together AI β€” meta-llama/Llama-3.3-70B-Instruct-Turbo, the former compiled default, is no longer served on serverless: every call answers Unable to access non-serverless model (2/9/1). The default moves to zai-org/GLM-5.3-Flash (11/0/1 on the last run of the day, 10/1/1 on the first with M10 red: the endpoint reports no cached tokens); Qwen/Qwen3.5-9B is 10/1/1 (M10). The default sees, so there is no separate vision companion any more.
  • HuggingFace β€” the default is 7/3/2 (M2, M5, M9) and Qwen/Qwen3.5-9B 9/1/2 (M2); the routing still decides how many of a model's providers carry tools.
  • OpenRouter and Mammouth AI β€” first campaigns, both on the fleet's Gemini default: one model, three transports. Direct 11/0/1, OpenRouter 11/1/0, Mammouth 10/0/2 β€” the marketplace returns the cache fields with 0 where Google returns none (M10), and Mammouth's two βž– are the capabilities the provider does not declare. OpenRouter's catalogue is the richest of the fleet: per model, the default parameters and whether reasoning is mandatory. Replayed on 2026-11-11: direct 11/0/1 (and gemini-3.8-flash 12/12), OpenRouter 12/0/0 on the default and 11/1/0 on gemini-3.8-flash (M10), Mammouth 9/1/2 on both β€” its M10 now reads 0 cached tokens where it was green on 2026-10-07.
  • Also served, not campaigned β€” among others claude-fable-5, claude-opus-4-8, gemini-3.6-flash, gemini-3.5-flash, mistral-large-2512, the ministral family, grok-4.5, grok-4.3, MiniMax-M2.1, MiniMax-M2.5, glm-5.1, glm-5-turbo, qwen3.6-plus, qwen3.5-plus. The snapshot lists them all.

Output caps β€” the documented maximum per model

A request carries an output cap (max_tokens, max_completion_tokens at OpenAI, num_predict at Ollama). Left unpinned, the cap is the model's documented maximum, from the LlmModelOutputLimits catalogue in Orkeon.Constants.Llm: a fixed 4096 would let a reasoning model spend the budget thinking and answer empty. A pinned value (Llm:MaxTokens, a profile, a crew's max_tokens) always wins; a model the catalogue does not know keeps the 4096 fallback. Only documented figures go in, each read on 2026-09-19:

Provider Model Cap sent Source Note
OpenAI (and Azure deployments of the same ids) gpt-5.6-sol, gpt-6-astra 128 000 developers.openai.com/api/docs/models reasoning counts inside the cap
Anthropic claude-sonnet-5, claude-opus-5, claude-fable-5-1 (dated variants follow the family) 128 000 platform.claude.com/docs/en/about-claude/models/overview thinking counts inside max_tokens; the field is required, an unknown Claude gets 4096
Gemini gemini-3.7-flash, gemini-3.8-flash (also under google/… on OpenRouter and bare on Mammouth) 65 536 ai.google.dev/gemini-api/docs/models includes thought tokens; 65 537 is a 400
DeepSeek deepseek-flash (and the retired deepseek-v4-flash* names it routes) 393 216 api-docs.deepseek.com/api/create-chat-completion "1 to 384K"; 384 000 on Mammouth
Kimi kimi-k3 131 072 platform.kimi.ai/docs/guide/kimi-k3-quickstart the vendor's own max_completion_tokens default; real bound 1M βˆ’ prompt
Kimi kimi-k3.1-pro 131 072 platform.kimi.ai/docs/guide/kimi-k3-quickstart a pin, not a documented cap: the K3 default carried over to the default model of 2026-11-11; kimi-k3 alone would not claim it (. is no separator), and the 4096 fallback starved K3
Kimi kimi-k2.6 131 072 platform.kimi.ai/docs/guide/troubleshooting a pin, not a documented cap: the bound is 256K βˆ’ prompt; a rejection drops the field on the retry
Qwen qwen3.7-plus, qwen3.8-flash, qwen3.8-max 131 072 alibabacloud.com/help/en/model-studio the chain of thought has its own thinking_budget; 65 500 on Mammouth
Z.AI glm-5.2, glm-5.3, glm-5.3-flash, glm-5.3-flashx 131 072 docs.z.ai/guides/overview/concept-param schema maximum; default 65 536
Z.AI glm-4.6v-flash 32 768 docs.z.ai/guides/vlm/glm-4.6v β€”
MiniMax MiniMax-M2 131 072 platform.minimax.io/docs/guides/models-intro "128k (including CoT)"; M3 is absent, default since 2026-11-11 β€” the vendor publishes no output figure, so it runs on the 4096 fallback (512k appears only as a benchmark setting; 512 000 on Mammouth)
xAI grok-4.6 128 000 docs.x.ai/developers/rest-api-reference no per-model cap; the vendor's own default when unset, reasoning excluded
Mistral mistral-medium-2604 (mistral-medium-3-5) none docs.mistral.ai/api/endpoint/chat no output cap, only "prompt + max_tokens ≀ context": the field is left out and the model writes to its window
Together any 4096 (fallback) docs.together.ai/docs/serverless-models no per-model output cap, only prompt + max_tokens ≀ window. The window itself was the cap from 2026-09-19 to 2026-09-21, on the assumption that context_length_exceeded_behavior: truncate clamps it to window βˆ’ prompt; the campaign of 2026-09-21 measured that two of three serverless engines refuse regardless (max_new_tokens, context_length_exceeded) and the third clamps on the buffered path only. The fallback holds, the flag is still sent, and the retry net drops the cap on those wordings; omitted, the field means 2048 (finish_reason: length)
Ollama any none (num_predict left out) docs.ollama.com/modelfile -1, infinite generation is the runtime's default: a local model writes to its context
HuggingFace, Docker Model Runner, anything else β€” 4096 (fallback) β€” the router's bound is the routed provider's context, which differs per route; pin Llm:MaxTokens

Two things the catalogue is honest about. Where the vendor bounds the cap by window βˆ’ prompt (Kimi, Together, Mistral), a fixed value near the window fails on any real prompt β€” so those rows are either a pin, the fallback, or nothing. And a cap the endpoint refuses (Qwen "Range of max_tokens", Kimi "prompt tokens + max_tokens exceeds", Gemini maxOutputTokens, DeepSeek's 422, Together's engines with max_new_tokens / context_length_exceeded) is retried once without the field when it came from the catalogue, with a warning naming the model β€” a pinned value is the user's, and its rejection surfaces unchanged.

Per-model mandatory parameter values

Some models refuse a request unless a parameter carries one specific value. This is a different animal from the per-model capability gaps above (a model lacking thinking or vision): here the capability exists, but the model dictates the value, and the vendor answers anything else with a 400. The constraint is per model, never per provider β€” the same vendor ships models with opposite demands, so a provider-level pin is one campaign away from breaking the sibling model. Every row below was measured live; nothing is inferred.

Provider Model Parameter Mandatory value Vendor's own words Measured
Kimi kimi-k2.6 temperature 1 invalid temperature: only 1 is allowed for this model 2026-08-03
Kimi kimi-k2.7-code temperature 1 β€” a buffered call is retried by the provider with that value, a stream is refused invalid temperature: only 1 is allowed for this model 2026-10-07
Kimi kimi-k3.1-pro, kimi-k3.1-pro-highspeed, kimi-k2.7-code-highspeed temperature 1 β€” every streaming mode and M2, M5, M9, M10 refused at 0; on the K3 family the provider now leaves the field out (with a warning) on every path invalid temperature: only 1 is allowed for this model 2026-11-11
OpenAI gpt-5.6-sol temperature 1 'temperature' does not support 0 with this model. Only the default (1) value is supported. 2026-08-30
OpenAI gpt-5.6-sol reasoning_effort "none" when the request carries function tools on /v1/chat/completions Function tools with reasoning_effort are not supported for gpt-5.6-sol in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'. 2026-08-30
OpenAI gpt-6-astra temperature 1 'temperature' does not support 0 with this model. Only the default (1) value is supported. 2026-09-21
OpenAI gpt-6-astra reasoning_effort no "none" exists (low, medium, high, xhigh) β€” the Sol workaround is impossible, so function tools stay refused on /v1/chat/completions until the dialect speaks /v1/responses 'reasoning_effort' does not support 'none' with this model. Supported values are: 'low', 'medium', 'high', and 'xhigh'. β€” and without it, Function tools with reasoning_effort are not supported for gpt-6-astra in /v1/chat/completions. To use function tools, use /v1/responses 2026-09-21
Anthropic claude-sonnet-5 temperature 1, or omit the field `temperature` is deprecated for this model. (1 and omission pass; 0 and 0.7 do not) 2026-08-30
OpenAI gpt-5.6-terra, gpt-5.6-luna, gpt-6-sol, gpt-6-luna temperature and reasoning_effort as gpt-5.6-sol: 1, and "none" with function tools the two sentences of gpt-5.6-sol, the model name aside 2026-10-07
OpenAI gpt-6.1-sol temperature and reasoning_effort as gpt-6-astra: 1, and no "none" exists 'reasoning_effort' does not support 'none' with this model. Supported values are: 'low', 'medium', 'high', and 'xhigh'. 2026-10-07
Anthropic claude-opus-5, claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5 temperature 1, or omit the field `temperature` is deprecated for this model. 2026-10-07
Anthropic claude-fable-5-2 temperature 1, or omit the field refused like the rest of the 5 family β€” the sentence was not recorded for this model 2026-11-11
OpenAI gpt-6.1-terra temperature and reasoning_effort as gpt-5.6-sol: 1, and "none" with function tools β€” unlike gpt-6.1-sol, this model has a none the two sentences of gpt-5.6-sol, the model name aside; 12/12 with those two pins 2026-11-11
Mistral mistral-large-4 reasoning_effort lifted: low, medium and high all accepted on 2026-11-13 (on 2026-10-07: high or none only) β€” 2026-11-13
Mistral mistral-small-2603 reasoning_effort high or none only β€” confirmed on 2026-11-13 reasoning_effort='low' is not supported for this model. Must be one of (<ReasoningEffort.none: 'none'>, <ReasoningEffort.high: 'high'>) 2026-11-13
Mistral mistral-medium-2604 reasoning_effort high or none only β€” confirmed on 2026-11-13 reasoning_effort low is not supported for this model, supported values: [<ReasoningEffort.high: 'high'>, <ReasoningEffort.none: 'none'>] 2026-11-13
Mistral mistral-medium-2604 top_p explicit 1 when temperature is 0 and reasoning is on (omission is NOT 1 there) top_p must be 1 when using greedy sampling. 2026-08-30
Anthropic claude-fable-5-1, claude-fable-5-2, claude-opus-5-5 thinking.type never disabled: thinking stays adaptive, the effort steers it "thinking.type.disabled" is not supported for this model. Use "thinking.type.adaptive" and "ou… (cut in the report) 2026-11-13
Anthropic claude-sonnet-5-5 thinking.type thinking is turned off with between_tools, not disabled To turn thinking off on this model, send "thinking": {"type": "between_tools"} instead of {"… (cut in the report) 2026-11-13
Kimi kimi-k2.7-code thinking.type enabled only: thinking cannot be turned off invalid thinking: only type=enabled is allowed for this model 2026-11-13
OpenRouter google/gemini-3.7-flash, google/gemini-3.8-flash reasoning.enabled never false Reasoning is mandatory for this endpoint and cannot be disabled. 2026-11-13
Z.AI glm-5.3, glm-5.3-flash, glm-5.3-flashx reasoning_effort low, high or max β€” no medium This model always engages in thinking and cannot be disabled; please use low, high, or max 2026-11-13

The rows dated 2026-11-13 come from the reasoning modes of the campaign (M15 thinking toggle, M16 effort sweep, M17 budget), which sent every level to every default and candidate model; the full grid, accepted values included, is Reasoning per model. Orkeon judges these values per model: the capability table (ModelCapabilityRegistry) narrows each provider's declaration, so for a model the table lists a refused off switch is dropped, a refused level is lowered to one the model takes (raised to the lowest when none lies below), and a refused vendor value (none, minimal) is dropped, each with a structured warning; for a model the table does not list, an effort the provider carries is sent as asked, with one warning per provider instance and model.

On OpenAI, the GPT-5.6 and GPT-6 tiers take function tools on /v1/chat/completions only with reasoning_effort: none (gpt-6-astra and gpt-6.1-sol have no none: tools need /v1/responses). The pin must reach the API as none: on 2026-11-11 a first pass had Orkeon lowering it to low, which turned M2 and M5 red on every OpenAI model until the fix.

A temperature nothing sets is not sent: a host that pins none runs these models on their own default, which they accept β€” orkeon init writes none. A host that pins one must pin the mandated value.

The counter-example that makes the registry per-model: gpt-4o-mini β€” same provider as gpt-5.6-sol β€” rejects reasoning_effort outright (Unrecognized request argument supplied: reasoning_effort, measured the same day). A provider-wide "always send none" would break it.

Machine-readable twin: llmproviders-test/lib/catalog.json, key requiredParams (per provider, keyed by model id) β€” the campaign scripts resolve it automatically and the report header prints the values actually used. In application code, express the same values through LlmConfig (Temperature, Thinking = { Effort = "none" }); on the wrong value the vendor's error comes back attributed (CapabilityMismatchHint names the thinking knob on the tools-with-reasoning refusal).

What this table does not prove

Every row is backed by unit tests asserting the emitted payload β€” against a mocked HTTP handler. A mock proves Orkeon sends what we believe it sends; it does not prove the vendor accepts it. Real-execution evidence is archived separately by the campaign kit (llmproviders-test/, whose README carries the campaign journal; the latest result per model is on LLM campaign results); run a campaign with orkeon llm probe --provider <name> --archive <dir>, or the kit's run-campaign.sh (validating a provider).