max-thinking is a single on/off scalar that decides whether reasoning-capable models think as hard as they're allowed to. on (the default) reproduces each model's full reasoning ceiling; off is the opt-in go-faster, spend-less mode that drops every reasoning model to its lightest supported level in one flip.
What it controls per model
Each verified reasoning model has a ceiling and a reduced level:
| Provider / model | on (ceiling) | off |
|---|---|---|
| Cerebras gpt-oss-120b | medium | low |
| OpenAI gpt-5.4 family | low | none |
| Groq / OpenRouter gpt-oss | low | low (their APIs reject none) |
| Cerebras gemma-4-31b, zai-glm-4.7 | non-reasoning or pinned; no effect | |
| Anthropic / Gemini | no reasoning knob; setting ignored | |
The design principle: on changes nothing versus how OpenCues always behaved. The ceilings were seeded to equal each provider's existing default reasoning effort, so on reproduces the prior behaviour exactly and off is the only state that alters anything. Turning it off is choosing speed and cost over depth on the surfaces where models reason: word cues, sentence cues, fluid-blank lookups, transform rewrites, and the background agent rewriter.
What it deliberately doesn't touch
The fluid-config classifier (the thing that turns stop showing tips _ into a settings change) always reasons at its own fixed low level, because its output is tiny and must be deterministic; a global speed knob shouldn't be able to make settings classification flakier. And explicit per-source reasoning floors still win, though they're clamped down to the model's ceiling, which keeps max-thinking a true cap rather than a suggestion.
Changing it
max-thinking: offin~/.cues/OPENCUES.md; hot-reload applies within about two seconds, so you can flip it mid-session and feel the difference immediately.- Cycle it in-text via
opencues settings _, like any other registry scalar. - With fluid config on, natural phrasings work:
think less _,turn off max thinking _.
Should you turn it off?
Honest guidance: if you run the default Cerebras path and your cues feel fast already, leave it on; the ceiling there is only medium, so there isn't much depth to trade away. The setting earns its keep when you're chasing latency on a reasoning-heavy provider, or minimising token spend on high-frequency surfaces like word cues that fire on ordinary typing. Note the quirky edges the docs call out: some providers' gpt-oss endpoints reject "no reasoning" outright with an HTTP 400 (their floor is low either way), one Cerebras model burns extra reasoning tokens at any level except none and so stays pinned there, and on non-reasoning models the toggle is simply inert rather than an error. The per-model ceiling table lives in packages/opencues-core/src/model-thinking.ts if you want the authoritative list.
Related: How OpenCues stays fast · How to switch models · What is OPENCUES.md? · GitHub repo