Reasoning effort

Each provider has its own dial: OpenAI reasoning_effort, Anthropic an effort config, Gemini a token-denominated thinking budget, GLM three fixed tiers, DeepSeek a model swap. Conifer takes the OpenAI-shaped field and translates per provider:

curl https://api.conifer.build/v1/chat/completions \
  -H "authorization: Bearer $CONIFER_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "glm-5.3",
    "max_tokens": 3000,
    "reasoning_effort": "medium",
    "messages": [{"role": "user", "content": "why is the sky blue?"}]
  }'

The levels#

none, minimal, low, medium, high, xhigh, max — the superset across providers. Few models serve all seven natively.

The matrix#

Probed 2026-08-29, one real request per cell against the live gateway, and re-probed after each snap shipped. ✓ = observed 200. Every cell is served except command-a-cohere, which has no effort dial at all and whose refusals relay honestly rather than being discarded.

A 422 cell is the upstream’s own refusal, relayed honestly rather than absorbed. After the snap rules below, the only refusing row left is command-a-cohere, which serves no effort dial at all; a — cell is one a rate limit kept the probe from measuring.

Where the gateway snaps for you#

OpenAI’s current reasoning models refuse minimal and max by name (“Supported values are: none, low, medium, high, and xhigh”), so the gateway snaps minimal to low and max to xhigh on OpenAI-proper seats — which is why every gpt row above reads as fully served.

GLM’s thinking models serve exactly three tiers — low, high, max — and 400 on every other spelling (“This model always engages in thinking and cannot be disabled”). An effort is a preference, not a contract term, so the gateway snaps the missing levels to the nearest served tier instead of relaying the refusal:

The snap is host-gated to Z.ai’s surface: no other provider sees the rewrite. If a GLM model refuses an effort level with “please use low, high, or max”, this shim is the fix — the refused spellings snap to the nearest tier the surface serves instead of erroring.

A third rule covers seats whose served band varies per model rather than per provider: Gemini 3.1 Pro’s floor is low while 3.5 Flash serves none, qwen3.8-27b tops out at medium, grok-4.5/4.6 refuse none and max, gpt-oss-120b starts at low. Each carries a measured floor-to-ceiling band and an out-of-band level clamps to the nearest end — which is why those rows above read as fully served.

Provider quirks#

When it still errors#

A refused level relays as 422 upstream_error carrying the provider’s message where it names the fix. 422, not 400: the request was well-formed on this wire; the provider declined it. Refused turns are not billed. A server fallback chain advances on these.