# Reasoning effort

> Control how long 7API models think. reasoning_effort, reasoning.effort, output_config.effort and thinking, plus the effort levels accepted by Claude, GPT, GLM, DeepSeek and Kimi models.


Most models in the catalog reason before they answer. Higher effort usually means better answers on hard problems, with more output tokens and a longer wait before the first visible token. Reasoning tokens are billed as output.

## Parameters

Send whichever parameter your client already uses. The gateway converts it to what the model expects.

| Parameter | Format | Example |
|---|---|---|
| `reasoning_effort` | Chat Completions | `"reasoning_effort": "high"` |
| `reasoning.effort` | Responses | `"reasoning": {"effort": "high"}` |
| `output_config.effort` | Anthropic Messages (used by Claude Code) | `"output_config": {"effort": "high"}` |
| `thinking` | Anthropic Messages | `{"type": "adaptive"}`. `{"type": "enabled", "budget_tokens": N}` is accepted on `claude-opus-4-6` only. |

## Levels by model family

| Models | Accepted levels | Notes |
|---|---|---|
| Claude (Opus, Sonnet, Haiku) | `low`, `medium`, `high`, `xhigh`, `max` | `claude-opus-4-6` has no `xhigh`. |
| GPT-6 / GPT-5.6 | `none`, `low`, `medium`, `high`, `xhigh`, `max` | `gpt-5.5` stops at `xhigh`. On `gpt-6.1-sol`, `max` becomes `xhigh` and `none` becomes `low`. |
| GLM 5.3, GLM 5.3 Flash | `low`, `high`, `xhigh`, `max` | Other values become `high`. |
| GLM 5.3 Abliterated | `high`, `xhigh`, `max` | The minimum and default is `high`. Thinking can't be turned off. |
| GLM 5.3 Uncensored, GLM 5.3 Flash Uncensored | `max` | Always maximum effort |
| DeepSeek V4 / V4.1 | `low`, `medium`, `high` | |
| DeepSeek V4.1 Flash 1017 Abliterated | `none`, `low`, `high`, `max` | Defaults to `high` |
| Kimi K3, Kimi K3.1 Pro | `low`, `high`, `max` | Defaults to `max`. `medium` becomes `high`, `xhigh` becomes `max`, and `none`, `off` and `minimal` become `low`. Any other value returns `400 invalid_reasoning_effort`. |

`GET /v1/models` reports `supported_reasoning_efforts` and `default_reasoning_effort` for models with a fixed scale. That lets clients such as OpenCode and Oh My Pi build their effort pickers automatically.

## Output budget

Reasoning and the final answer share the request's output limit (`max_tokens` or an equivalent). At high effort, set the limit generously. The GLM abliterated models default to a 130,000-token output budget for this reason.
