Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.clearmaas.com/llms.txt

Use this file to discover all available pages before exploring further.

Reasoning models spend extra compute on a hidden “thinking” pass before producing the final answer. They’re slower and more expensive but solve harder problems. ClearMaas provides one unified syntax for controlling reasoning effort across every provider — pick whichever form fits your client.

​
Two ways to set effort

​
1. The reasoning_effort field (OpenAI shape)

Pass it on a Chat Completions request. Values: low, medium, high (and minimal / max on some models). 
resp = client.chat.completions.create(
    model="anthropic/claude-opus-4.6",
    messages=[{"role": "user", "content": "Hard math problem..."}],
    reasoning_effort="high",
)
ClearMaas translates this field to the upstream’s native shape:
  • OpenAI o-series and gpt-5-pro family: forwarded as native reasoning_effort.
  • Anthropic Claude: mapped to thinking: {type: "enabled", budget_tokens: ...} with budgets low→1280, medium→2048, high→4096. For claude-opus-4.6 specifically, mapped to thinking: {type: "adaptive"} plus output_config.effort.
  • Google Gemini: mapped to generationConfig.thinkingConfig with includeThoughts: true and a thinking-level / budget set from the effort.
  • xAI Grok: forwarded for grok-3-mini family (which accepts reasoning_effort natively).
  • DeepSeek reasoner: model is reasoner-by-design; reasoning_effort is a no-op.

​
2. The -{effort} model-name suffix

You can also bake the effort into the model name. Recognized suffixes: -minimal / -low / -medium / -high / -max. 
# Equivalent to model="anthropic/claude-opus-4.6" + reasoning_effort="high"
resp = client.chat.completions.create(
    model="anthropic/claude-opus-4.6-high",
    messages=[...],
)
Works the same way across providers — pick whichever line is more readable in your code.

​
Reasoning model families in this deployment

OpenAI:
  • openai/o1, o1-pro
  • openai/o3, o3-mini, o3-mini-high
  • openai/o4-mini, o4-mini-high
  • openai/gpt-5-pro and gpt-5.x-pro family
Anthropic (extended thinking on Claude 4 / Opus):
  • anthropic/claude-sonnet-4.6, claude-opus-4.6, claude-opus-4.7, etc. — pair with reasoning_effort or the -{effort} suffix.
Google Gemini (extended thinking on Gemini 2.5 / 3.x):
  • google/gemini-2.5-pro, gemini-2.5-flash, gemini-3-pro-preview, etc. — pair with reasoning_effort or the -{effort} suffix.
DeepSeek:
  • deepseek/deepseek-reasoner — reasoner-by-design.
xAI Grok:
  • grok/grok-4-fast-reasoning, grok-4-1-fast-reasoning
  • grok/grok-3-mini paired with reasoning_effort: low or high
Call /v1/models for the live catalog.

​
Reasoning trace in the response

For OpenAI Responses API the model’s hidden reasoning is returned as reasoning items in the response output. For Anthropic via native /v1/messages, thinking arrives as content_block entries of type thinking. The gateway also surfaces a reasoning_content field on chat-completion responses where the upstream provides one. You can display the trace for transparency or ignore it in production.

​
Billing

Reasoning tokens are tracked separately on completion_tokens_details .reasoning_tokens in the response usage object — see Operations / Billing & Usage.