AnyRouter should unify reasoning effort across upstreams and models

Open ticket for project-anyrouter, measured on the full catalog (GET /api/v1/models?all=1, 206 listings, 2026-09-27). AnyRouter is one gateway over many upstreams, so it should expose one reasoning contract instead of whatever each upstream happens to accept.

Evidence:

• 123 listings carry reasoning in capabilities, but only 12 list any reasoning request field in supported_parameters.
• Those 12 disagree with each other: inclusionai/ling-3.0-flash-fin and nex-agi/nex-n2.5-* list only reasoning; z-ai/glm-4.5 and z-ai/glm-5.2 list only reasoning_effort; only z-ai/glm-5.3-flash and stealth/space-bunny-alpha list all three.
• No listing ever publishes the accepted effort values. A client cannot tell whether xhigh is legal on a given route.
• All 9 anyrouter/* presets advertise only [max_tokens, temperature, top_p, stop, stream] — no reasoning and no tools — even though auto/hermes/cowork route to reasoning models and the preset descriptions tell callers to pass tools.
• 101 of 206 listings have top_provider.max_completion_tokens: null, so the documented clamp ("A max_tokens set higher is clamped down silently") is undiscoverable. anyrouter/decision is one of them.

Asks:

1. One reasoning surface for the whole API: the same request field (reasoning + reasoning_effort) accepted on every model, forwarded to whichever route serves the request.
2. Publish the accepted effort values per model, or adopt a single ladder (low/medium/high) and document it once.
3. Include reasoning and tools in the presets' supported_parameters.
4. Fill top_provider.max_completion_tokens (or add an explicit max_output_tokens) so the clamp is discoverable.

The contract is already documented — chat is effort-only

/api-reference/models states it outright, and it settles the toggle question that no amount of probing could:

"AnyRouter does not infer universal toggle, token-budget, or interleaved-thinking support from a model name or description. Toggle and budget controls remain endpoint-specific (reasoning.enabled / reasoning.max_tokens on Responses, thinking on Messages)."

The architecture doc in the gateway repo (docs/kb/architecture/reasoning-metadata-and-preset-contract.md) adds: for anyrouter/* presets, "Chat uses flat reasoning_effort; Responses uses nested reasoning.effort plus endpoint-specific reasoning.enabled; Messages uses thinking; none remains a disable sentinel on the flat Chat/converted-Messages path." So on chat there is one control plus an off sentinel, and a toggle/budget_tokens entry in any catalog for this API is wrong.

Probes agree (2026-09-27, POST /api/v1/chat/completions, reasoning object forms effort / enabled / max_tokens / exclude):

route	reasoning: {…}	reasoning_effort: none
deepseek/deepseek-v4.1-flash	400 Validation: Unsupported parameter(s): 'reasoning'	200, reasoning_tokens: 0
openai/gpt-oss-20b	400 property 'reasoning' is unsupported	400 (upstream allowlist is low/medium/high only)
nvidia/nemotron-3-ultra-550b-a55b	200; enabled:false → 0, exclude:true ignored	200, 0
anyrouter/free	200; enabled:false → 0	200, 0

The nemotron routes accept the object form only because they are served by an OpenRouter-compatible pool that passes it through — exactly the "provider support remains route-specific" caveat in the host's own doc. It is not a gateway contract, so catalogs must not publish a toggle on it.

Measured: the gateway forwards the field and never validates it

Reading error.metadata.upstream_message, the serving upstream's 400 decides which rungs are legal:

upstream	its own allowlist
openrouter-pool	max, xhigh, high, medium, low, minimal, none
huggingface-pool	low, medium, high — rejects xhigh
nvidia-byok (deepseek-v4.1-flash)	rejects minimal and medium
z-ai-coding-byok	no validation, accepts any value
hermes-agent	low, medium, high, xhigh

The accepted set is per model id, not per backend: nvidia-byok serves both deepseek-v4.1-flash (rejects minimal/medium) and muse-glimmer-30b (accepts both). A published low/medium/high contract returns a hard 400 on deepseek-v4.1-flash.

Honored effect, same multi-step prompt, mean reasoning_tokens (2026-09-27):

route	none	low	medium	high	max
openai/gpt-oss-20b (700)	rejected	273	629	698	rejected
deepseek-v4.1-flash (1500, n=3)	0	718	rejected	1182	1410
nemotron-3-super-120b-a12b (1500, n=3)	0	663	—	916	722
nemotron-3-ultra-550b-a55b (1500, n=3)	0	615	—	471	287
anyrouter/free (700)	0	535	580	626	—
minimax/m3 (700)	0 / 349	611	551	846	—

none is a real off position everywhere it is accepted. nemotron-3-ultra is the counter-example to "deeper value = more tokens": its ladder came out inverted, so accept-without-honored-effect is the honest reading for that route — publish the peer's rungs, not a wider set.

Presets resolve the member per request, so their ladder is not separable: cowork/hermes served poolside/laguna-s-2.1 (none = 0, every graded value pinned at the whole budget), auto served nvidia/nemotron-3-ultra-550b-a55b (low = 3000), latest served meta/muse-glimmer-30b (no reasoning tokens reported at all), free served dots-studio/dots-3-note-preview.

Account-dependent gotchas when re-probing: minimax/m3 ran out of upstream credit partway through the day (402 → 502, not a rejected rung), and z-ai/glm-4.7-flash is intermittent (502 from z-ai-coding-byok). Eight ids are BYOK-only and cannot be exercised with a plain key at all: 404 Model … is a Bring Your Own Key (BYOK) only model or No upstream available for model … on every value, including a nonsense one.

none is not uniformly an off switch. Two routes show why one measurement is not enough:

• minimax/m3, five runs of none at max_tokens=1200: the two the route served returned 0 and 787 reasoning tokens. Same value, same route, so the sentinel can be ignored there and no catalog should promise it.
• anyrouter/hermes answers none with 400 Reasoning is mandatory for this endpoint and cannot be disabled. while low/medium/high each return 0 tokens on the member it does serve — publishing the off switch on that preset would advertise a 400.
• By contrast anyrouter/auto, cowork and free return 0 on every served run of none (5/5, 3/3, 4/4) and spend the whole budget on any graded value, because the router picks the member per request: a preset can promise the off and nothing deeper.

Why: without this, every client maintains its own per-model reasoning table and guesses, and a uniform contract silently breaks on the routes whose upstream is narrower. How to apply: no public tracker exists for AnyRouter (closed service, no GitHub repo, issues disabled here), so this note is the ticket of record. Ask #2 (publish the accepted values per listing) is the one that unblocks full-fidelity per-model data; until then publish the route's measured set, cite the measurement in the entry, drop every toggle and budget_tokens, and use none as the off position, and only where it reproduces. Shipped that way in the models.dev catalog (PR anomalyco/models.dev#8023, merged after the automated review came back clean): per-model ladders, no toggles, no cost tiers (this API publishes flat rates), no per-token price for the anyrouter/* presets because they bill the route they resolve, and [] for the five entries where nothing could be verified — m3, kimi-k2.6, grok-build-0.1, hermes, latest. See project-anyrouter-openai-compat.