Reasoning-capable models can think before they answer: better quality, slower and more expensive. max-thinking is one switch that decides whether each model thinks as hard as it is allowed to, or drops to a faster, lighter level.
| max-thinking | Effect |
|---|---|
| on | the default: each model reasons up to its own tested ceiling |
| off | faster and cheaper: each model thinks minimally, or not at all where the model allows it |
on reproduces the behaviour OpenCues has always had. off is the opt-in "go faster" mode.
Every verified model has a tested ceiling and a reduced level.
| Provider and model | on | off |
|---|---|---|
| Cerebras gpt-oss-120b | medium | low |
| Groq gpt-oss-120b / 20b | low | low |
| OpenAI gpt-5.4 family | low | none |
| OpenRouter gpt-oss | low | low |
| Cerebras zai-glm-4.7, gemma-4-31b | none | none |
| Anthropic, Gemini | no reasoning knob; the setting has no effect | |
Where on and off read the same, that is the model's floor: some providers reject a request to disable reasoning outright, so low is as low as they go.
Every prose and lookup surface: word alternatives, sentence rewrites, _ fills, transforms, and agent rewrites. The settings classifier is the one exception; its output is tiny and must be dependable, so it pins low regardless of the switch, capped at whatever its model allows.
| Where | How |
|---|---|
| the file | max-thinking: off in ~/.cues/OPENCUES.md; picked up within about two seconds |
| the settings menu | opencues settings _, cycle to max-thinking, Ctrl+Alt+Up/Down between on and off |
| plain words | type think less _; the settings classifier routes it |