What is reasoning effort? Low, medium, high and max

Anthropic (Claude)

• Set output_config.effort. Effort affects all output tokens, including text, tool calls and thinking, per Anthropic's effort docs.
• low is the most efficient. xhigh is for long-running agentic and coding tasks of over 30 minutes. max has no constraint on token spending.
• API defaults: medium on Opus 5.5 and Haiku 5.5, high on other models, including Sonnet 5.5 and Fable 5.1.
• Claude Code starts Opus 5.5, Sonnet 5.5 and Haiku 5.5 at medium. Change it with /effort or --effort.
• Effort also affects how fast you use your Claude plan's limits.

OpenAI (GPT-6 and Codex)

• Set reasoning.effort in the Responses API. Supported values depend on the model: GPT-6 Astra and GPT-6.1 Sol accept low to max, while GPT-6 Sol and GPT-6 Luna also accept none.
• GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna default to medium. OpenAI calls medium a well-balanced default for most workloads.
• In the Codex apps the levels are called Light, Medium, High and Extra High, plus Max. Ultra, separately, splits work across subagents.
• In the CLI, use /model or set model_reasoning_effort in config.toml. OpenAI suggests starting at High for Luna and Light for Astra.

Choosing a level

Both companies say to start at the default and measure. Anthropic notes that tuning effort is often a better lever than switching models, and OpenAI says levels don't map exactly between model generations. Higher effort costs more tokens, so on a subscription it also uses your limits faster.