Claude Opus 4.6 Release: Adaptive Thinking and API Changes
Anthropic released Claude Opus 4.6 on February 5, 2026. This update introduces adaptive thinking, deprecates manual budgets, and removes fast mode support.
Official Announcement and Core Specifications
Anthropic officially launched Claude Opus 4.6 on February 5, 2026. The provider describes this model as its most intelligent option for complex agentic tasks and long-horizon work. The model supports text and image inputs while producing text outputs. It features a context window of 1,000,000 tokens and a maximum output limit of 128,000 tokens. Open weights are not confirmed for this release.
Thinking Mechanisms and Mode Restrictions
The release notes specify that Opus 4.6 recommends adaptive thinking using the configuration `thinking: {type: "adaptive"}`. Manual thinking, defined as `type: "enabled"` with `budgettokens`, is deprecated for this model. Additionally, the model does not support prefilling assistant messages. A separate changelog entry states that fast mode has been removed for Claude Opus 4.6. Requests specifying `speed: "fast"` now run at standard speed and are billed at standard rates without returning an error. The response's `usage.speed` field reports the actual speed used.
- Adaptive thinking is the recommended configuration.
- Manual thinking with budget tokens is deprecated.
- Assistant message prefilling is unsupported.
- Fast mode requests default to standard speed and pricing.
Evaluation Procedure for Integration
Engineers can evaluate the model by testing its response to specific API parameters. First, send a request using the adaptive thinking configuration and measure the resulting token usage. Next, test a request with the deprecated manual thinking parameter to validate how the system handles the deprecation. Finally, compare the billing rate and speed field when submitting a request with `speed: "fast"` against a standard request. These steps verify the documented behavior changes without assuming performance improvements.
Artificial Analysis benchmark review
Benchmark review updated on . These scores apply to the linked artifact and recorded methodology; they do not establish the performance of a different runtime or quantized model.
| Measurement | Recorded result | Measured |
|---|---|---|
| Artificial Analysis Intelligence Index (v4.1) ↗ | 38.8 index_points | 2026-08-09 |
| Artificial Analysis Intelligence Index (v4.2) ↗ | 30.8 index_points | 2026-09-05 |
| Artificial Analysis Intelligence Index (v4.3) ↗ | 26.4 index_points | 2026-09-09 |
Review the full benchmark provenance and source methodology →
Primary sources
Frequently asked questions
Does Claude Opus 4.6 support fast mode?
No. The source confirms that fast mode was removed for this model. Requests with `speed: "fast"` execute at standard speed and incur standard billing rates.
Is manual thinking still supported?
Manual thinking is deprecated. The documentation recommends using adaptive thinking instead. The source does not state whether manual thinking will cause errors or simply operate differently.