Models & Thinking Modes

Current models

Standard API rates verified October 27, 2026. USD per million input/output tokens, excluding caching.

Model	Input / output	Context	Default effort
Haiku 5.5	$0.10 / $0.50 (over 100K: $0.50 / $2.50)	1M	medium
Sonnet 5.5	$2 / $10	1M	medium
Opus 5.5	$4 / $20	1M	medium
Fable 5.1	$10 / $50	1M	high

Sonnet’s published rate remains $2/$10 after the launch promotion. Opus 5.5 is the direct-service default; account policy can override it. Model choice does not establish a typical task cost. Effort shown is the Claude Code default; the API default for Sonnet 5.5 is high.

Select and save

/model opens the picker. Enter saves a default; s applies it to this session only. claude --model sonnet applies to the launched session. Project and managed settings can reapply at the next launch.

Aliases depend on the provider. opus maps to 5.5 on most providers, but 4.6 on Foundry. sonnet maps to 5.5 on the direct API, 4.6 on Claude Platform on AWS, and 4.5 on Bedrock, Google, and Foundry.

haiku selects 5.5 on the Anthropic API (v2.1.293+), 4.5 on AWS, Bedrock, Google, and Foundry. fable selects 5.1 where available, or 5 through the Claude apps gateway. best selects Fable when available, otherwise Opus. Pin a supported full ID when version identity matters.

Thinking and effort

Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1 use adaptive thinking. Legacy Haiku 4.5 has no effort parameter.

Current adaptive models support low, medium, high, xhigh, and max. Opus 4.6 and Sonnet 4.6 support max but not xhigh.

/effort saves levels through xhigh per model. Use s for a session-only choice. max is session-only; /effort auto clears the active model’s saved choice.

Option+T / Alt+T toggles thinking where supported. It has no effect on Opus 5.5, Sonnet 5.5, Haiku 5.5, or Fable, whose thinking cannot be turned off. Lower effort for bounded work instead.

ultrathink adds a reasoning instruction without changing API effort. think hard is ordinary prompt text.

Choose by measured results

Evaluate Haiku on bounded tasks, Sonnet on routine work, and Opus or Fable on harder tasks. Compare accepted outcomes, retries, latency, and total cost on the same task set. Teammates can use different models.

Sources: Model configuration, API pricing.