Claude Opus 4.6

The Opus Generation Where the Reasoning Contract Started to Change

Anthropic released Claude Opus 4.6 on February 5, 2026. It introduced adaptive thinking and generally available effort controls while still accepting the older manual extended-thinking format based on budget_tokens.

That makes Opus 4.6 different from both earlier and later generations. Earlier Claude 4 models relied on manual extended thinking. Opus 4.7 and newer Opus models removed that legacy mode and require adaptive thinking when reasoning is enabled.

Opus 4.6 also helped establish the modern large-context Opus baseline. It now provides a 1,000,000-token context window at standard pricing, supports up to 128,000 standard output tokens, accepts text and images, and supports tool-oriented workflows.

Anthropic lists the model as Active (legacy) and recommends migrating to Claude Opus 5.5 for improved performance.

• Model ID: claude-opus-4-6.
• Released February 5, 2026.
• Active legacy model.
• 1M context window.
• 128K standard maximum output.
• Text and image input with text output.
• Adaptive thinking supported; legacy extended thinking still accepted but deprecated.

Judge the Entire Execution Path

Long-horizon agents operate differently from ordinary chat. A coding agent may inspect files, call tools, modify code, run tests, investigate failures, update its plan, and continue for many turns before producing a final answer.

The same pattern appears outside software engineering. A research agent may retrieve evidence repeatedly, a data workflow may call specialized tools, and an enterprise agent may maintain instructions across a long chain of actions.

For these workloads, the useful unit of evaluation is not one response. Measure the complete trajectory:

• Did the model finish the task?
• How many tool calls were necessary?
• Did it repeat work?
• Did it recover from errors?
• How much context accumulated?
• How many tokens were spent on reasoning and output?
• How often did a human need to intervene?

Opus 4.6’s 1M context window and tool support make it suitable for this style of workload, but later Opus versions improve both model capability and integration behavior. Historical production data from 4.6 is therefore valuable as a regression baseline.

Agentic Cost Depends on More Than the Headline Rate

Anthropic lists 5-minute cache writes at $6.25 per million tokens, 1-hour cache writes at $10 per million, and cache reads at $0.50 per million.

These rates matter for long-running workflows because the same large context may be reused repeatedly. System instructions, repository snapshots, reference documents, tool definitions, or policy material can often remain stable across multiple turns.

Batch API requests receive a 50% discount on standard input and output pricing. This makes batch execution useful for large offline evaluations, bulk code or document processing, and workloads that do not need interactive latency.

Current Claude Opus 5.5 is cheaper at $4 input and $20 output per million tokens. That means migration from Opus 4.6 can lower the base token rate even before quality improvements are considered.

• Input: $5.00 per 1M tokens.
• Output: $25.00 per 1M tokens.
• 5-minute cache write: $6.25 per 1M tokens.
• 1-hour cache write: $10.00 per 1M tokens.
• Cache read: $0.50 per 1M tokens.
• Batch API: 50% discount on input and output.

No Assistant Prefill, Legacy Tool Semantics, and Historical Fast Mode

First, Opus 4.6 does not support prefilling assistant messages. Applications that previously started an assistant response with predefined text need to use a different output-control pattern.

Second, Opus 4.6 still accepts forced tool choice and the older computer_20251124 computer-use tool. These behaviors change in newer Opus generations.

Third, Anthropic launched a research-preview fast mode for Opus 4.6 in February 2026, with output generation up to 2.5× faster at premium pricing. That fast mode was removed on June 29, 2026.

Today, sending speed: "fast" to Opus 4.6 does not provide fast inference or premium billing. The request runs at standard speed and standard rates, and usage.speed reports the actual speed used.

Finally, the 4.6 generation introduced dateless model IDs. claude-opus-4-6 identifies a pinned model version without the snapshot date format used by earlier Claude releases.

On Amazon Bedrock, Opus 4.6 is also a boundary case: it is the last model whose Bedrock identifier keeps the -v1 suffix.

A Feature-Rich Compatibility Baseline

The model accepts text and images and produces text. Adaptive thinking is the recommended reasoning path, but legacy extended thinking remains available for older integrations.

Tool use supports complex agentic workflows, and prompt caching reduces repeated processing cost for stable context. The model also supports large Batch API outputs and was the launch model for Anthropic’s compaction API.

Its integration surface is more permissive than current Opus in several ways: thinking can be absent or explicitly disabled, forced tool choice is available, and the older computer-use tool is accepted.

At the same time, some newer features are absent or behave differently. Assistant prefilling is unsupported, high-resolution vision arrived in Opus 4.7, and later models introduce additional dynamic orchestration controls.