The first Opus-class model with a million-token window, built for agents that run across whole workflows rather than single prompts. Try it here, then open the full studio.
Opus 4.6 was Anthropic's February 2026 flagship, focused on three things: reasoning depth via adaptive thinking, a million-token context, and agentic task execution.
Published figures, including the rows where Opus 4.6 isn't ahead. Higher is better; a dash means no directly comparable published figure.
| Benchmark | Claude Opus 4.6 | GPT-5.2 | Gemini 3 Pro |
|---|---|---|---|
| SWE-bench Verified· real GitHub issues | 80.8% | 80.0% | 76.2% |
| Terminal-Bench 2.0· command-line agents | 65.4% | — | — |
| OSWorld· agentic computer use | 72.7% | — | — |
| τ²-bench Retail· tool orchestration | 91.9% | — | — |
| BrowseComp· agentic browsing | 84.0% | — | — |
| ARC-AGI-1· abstract reasoning | 94.0% | 90%+ | — |
| ARC-AGI-2· harder abstract reasoning | 69.2% | — | 31.1% |
| MRCR v2 8-needle 1M· long-context recall | 76.0% | — | — |
| GPQA Diamond· graduate science | — | 93.2% | 91.9% |
| Context window· input tokens | 1M | 400K | 1M |
Figures from Anthropic's Opus 4.6 announcement and system card. The honest picture: GPT-5.2 leads on GPQA Diamond, and some users reported Opus 4.6 writes flatter prose than Opus 4.5.
The first Opus-class model with a million-token window, so an agent can hold an entire codebase or document set without losing the thread partway through.
Four effort levels — low, medium, high and max — let the model decide when deeper reasoning actually helps, rather than burning tokens on easy questions.
80.8% on SWE-bench Verified and 65.4% on Terminal-Bench 2.0. Built for large refactors and multi-step debugging that unfolds over hours, not single-file edits.
91.9% on τ²-bench Retail. It holds up when coordinating many tools at once, which is where most agent setups fall apart.
72.7% on OSWorld — strong enough to drive a desktop through multi-step workflows rather than just describing what to click.
During pre-release testing Anthropic reported it independently surfaced over 500 previously unknown zero-day vulnerabilities in open-source code.
The model is identical. What differs is everything around it.
Opus 4.6 was built around a specific bet: that the valuable thing is no longer answering one prompt well, but staying coherent across a task that runs for hours. Everything about it follows from that.
This is where it's strongest. Large refactors, migrations, multi-step debugging — the kind of work where a model that loses the plot halfway through is worse than useless. The million-token context and 76% recall on the 8-needle 1M MRCR test mean it can actually hold a large codebase in mind.
91.9% on τ²-bench Retail and 72.7% on OSWorld put it among the strongest options for orchestrating tools and driving a computer. If your agent breaks down when it has more than a handful of tools available, this is the model to test against.
Opus 4.6 replaced extended thinking with four effort levels: low, medium, high and max, with high as the default. The model decides how much reasoning a given question deserves, so easy prompts finish early instead of burning tokens.
Two honest caveats. First, GPT-5.2 scores higher on GPQA Diamond (93.2%), so for pure scientific Q&A it may serve you better. Second: a meaningful number of users found Opus 4.6 produced flatter, more generic prose than Opus 4.5. If writing is your main use, test it against alternatives on your own material before switching wholesale.