Thinking levels
A model that reasons thinks before it answers. How much it thinks is set with one level, the same on every route and every model: OpenAI's reasoning_effort values. HIPS translates the level into what the model you call takes, so you set it once and never need to know each model's own words for it.
The central levels, lowest first
| Level | Meaning |
|---|---|
none | No thinking, where the model can turn it off. |
minimal | As little as the model allows. |
low, medium, high | Less or more thinking; medium is the usual default. |
xhigh, max | The most the model allows. |
- Level
none- Meaning
- No thinking, where the model can turn it off.
- Level
minimal- Meaning
- As little as the model allows.
- Level
low,medium,high- Meaning
- Less or more thinking;
mediumis the usual default.
- Level
xhigh,max- Meaning
- The most the model allows.
Every level is accepted on every model, so a valid level is never refused. On a model that does not reason it is accepted and has no effect. A value outside the seven is refused with a 400 invalid_value that names the field.
| Route | How to set it |
|---|---|
| Chat completions | reasoning_effort, one of the seven levels. |
| Responses | reasoning.effort, one of the seven levels. |
| Messages | thinking and output_config.effort. thinking: {"type": "disabled"} is none. thinking: {"type": "enabled", "budget_tokens": N} is low for a budget under 4,096 tokens, medium under 16,384 and high above. adaptive, or enabled with no budget, leaves the model at its own level. output_config.effort (low, medium, high or max) is the level itself, and is used over a budget; a disabled thinking is used over both. |
- Route
- Chat completions
- How to set it
reasoning_effort, one of the seven levels.
- Route
- Responses
- How to set it
reasoning.effort, one of the seven levels.
- Route
- Messages
- How to set it
thinkingandoutput_config.effort.thinking: {"type": "disabled"}isnone.thinking: {"type": "enabled", "budget_tokens": N}islowfor a budget under 4,096 tokens,mediumunder 16,384 andhighabove.adaptive, orenabledwith no budget, leaves the model at its own level.output_config.effort(low,medium,highormax) is the level itself, and is used over a budget; a disabled thinking is used over both.
Each model's sheet in the catalog says, under Reasoning, which of these kinds it is. A model with no Reasoning section does not reason, and the level has no effect on it.
| Kind | What the level does |
|---|---|
| Levels | The model's chat template takes named levels. The level you send goes as the nearest level the model takes; between two equally near, the lower one. A model that takes only low, medium and high is sent high for xhigh or max and low for minimal. none turns thinking off where the model can be turned off; where it cannot, it goes as the model's lowest level, the closest it comes. A level the model has no word for, and that is nearer its own default, leaves it at its default. A model that refuses some levels is never sent one it refuses. |
| On or off | none turns thinking off. Any other level leaves it on, nothing finer. |
| Its own level | The model thinks at its own level. The level is accepted, and nothing of it is promised. |
- Kind
- Levels
- What the level does
- The model's chat template takes named levels. The level you send goes as the nearest level the model takes; between two equally near, the lower one. A model that takes only
low,mediumandhighis senthighforxhighormaxandlowforminimal.noneturns thinking off where the model can be turned off; where it cannot, it goes as the model's lowest level, the closest it comes. A level the model has no word for, and that is nearer its own default, leaves it at its default. A model that refuses some levels is never sent one it refuses.
- Kind
- On or off
- What the level does
noneturns thinking off. Any other level leaves it on, nothing finer.
- Kind
- Its own level
- What the level does
- The model thinks at its own level. The level is accepted, and nothing of it is promised.
On HIPS's data-center capacity in the Certified pool, the level is passed in the form each model there takes it, the nearest it takes; where a model there takes no level, it is accepted and has no effect.
The tokens a model spends thinking are output tokens, billed as completion_tokens (output_tokens on Messages and Responses), whether or not the route returns the thinking: Messages returns none of it. Turning thinking off with none, where the model can turn it off, is how to stop paying for it.
Note
A level is never sent to a model in a form it does not take: it reaches each model as the nearest level that model takes, or not at all.
Related sheets
- Sheet
- 20 of 21
- Detail
- F
- Updated
- 2026-11-18