Thinking and reasoning effort

Both abliterated-model and abliterated-model-large are reasoning models: they think before answering by default. You can tune how much they think, disable it, or keep the reasoning but hide it from the response.

Effort levels

The reasoning depth is set with an effort value. From least to most: minimal · low · medium · high · xhigh · max none disables reasoning entirely.

Model-aware effort. abliterated-model honors distinct levels on Chat Completions. abliterated-model-large runs in two reasoning modes — high and max — so it accepts every value but maps them: minimal–high → high, and xhigh–max → max. Accepting several aliases does not mean the large model has six distinct depths.

Setting effort per endpoint

The field differs by API surface. Chat Completions (/v1/chat/completions) — set reasoning_effort. Responses (/v1/responses) — set reasoning.effort. Anthropic Messages (/v1/messages) — set output_config.effort, or request extended thinking with thinking.budget_tokens. If both are given, output_config.effort wins.

On Responses, the base model accepts none, minimal, low, medium, high, and xhigh — but not max. Only abliterated-model-large accepts max. On Chat Completions, both models accept the full ladder.

A top-level effort also works on Chat Completions and Responses as a non-standard alias (not on Messages). Standard OpenAI-compatible clients should prefer reasoning_effort or nested reasoning.effort.

Where the reasoning trace appears

Endpoint	Reasoning trace location
Chat Completions	`choices[].message.reasoning_content`
Responses	a `reasoning` item in `output[]`
Anthropic Messages	a `thinking` content block

Disabling vs. hiding reasoning

These are different, and they affect token usage differently.

Disable — the model does not reason; only the final answer is generated (fewer tokens):
- Chat Completions / Responses: reasoning_effort: "none", reasoning.effort: "none", or effort: "none"
- Anthropic Messages: thinking: false
- Legacy top-level thinking: false also works on Chat Completions.
Hide — the model still reasons (tokens are still spent) but the trace is omitted from the response:
- Chat Completions and Responses: include_reasoning: false (or reasoning.exclude: true)
- Anthropic Messages: include_reasoning: false removes the thinking blocks from the response; thinking.display: "omitted" keeps a thinking block but empties its text.

include_reasoning: false does not save reasoning tokens — it only removes the trace from the response. To reduce token use, disable reasoning with "none".

Defaults

Reasoning is on by default when no control is supplied.

Model	Default
`abliterated-model`	Reasoning on, at the model’s native depth
`abliterated-model-large`	Reasoning on at high

Legacy thinking: true just means “reasoning on” — it is not a specific effort level.

Examples

Chat Completions

curl -s https://api.abliteration.ai/v1/chat/completions \
  -H "Authorization: Bearer $ABLIT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "abliterated-model-large",
    "reasoning_effort": "max",
    "messages": [{ "role": "user", "content": "Prove that sqrt(2) is irrational." }]
  }'

The base model accepts the same field with any level, e.g. "reasoning_effort": "medium" on "model": "abliterated-model".

Responses

curl -s https://api.abliteration.ai/v1/responses \
  -H "Authorization: Bearer $ABLIT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "abliterated-model",
    "reasoning": { "effort": "high" },
    "input": "Prove that sqrt(2) is irrational."
  }'

For literal max on Responses, use abliterated-model-large — the base model accepts up to xhigh but rejects max there.

Anthropic Messages

curl -s https://api.abliteration.ai/v1/messages \
  -H "x-api-key: $ABLIT_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "abliterated-model",
    "max_tokens": 1024,
    "output_config": { "effort": "high" },
    "messages": [{ "role": "user", "content": "Prove that sqrt(2) is irrational." }]
  }'

You can also request extended thinking with a token budget: "thinking": { "type": "enabled", "budget_tokens": 512 }. The budget is mapped to an effort level; if you also set output_config.effort, that takes precedence.

Get started

Capabilities

Integrations

Policy Gateway

Reference

Thinking and reasoning effort

Effort levels

Setting effort per endpoint

Where the reasoning trace appears

Disabling vs. hiding reasoning

Defaults

Examples

Chat Completions

Responses

Anthropic Messages

​Effort levels

​Setting effort per endpoint

​Where the reasoning trace appears

​Disabling vs. hiding reasoning

​Defaults

​Examples

​Chat Completions

​Responses

​Anthropic Messages

Effort levels

Setting effort per endpoint

Where the reasoning trace appears

Disabling vs. hiding reasoning

Defaults

Examples

Chat Completions

Responses

Anthropic Messages