Claude Opus 4.6: Anthropic's 1M Context Window Changes Everything

Anthropic released Claude Opus 4.6 on February 5, 2026, and the headline spec is a 1 million token context window paired with 128K output tokens. These aren’t incremental bumps, they change the category of work you can hand a model in one go. (Anthropic has since moved on: Opus 4.7 followed on April 16 and Opus 4.8 became the flagship on May 28, 2026. The analysis below reflects Opus 4.6 as it stood at launch.)

The Context Window Number in Practice

One million tokens is roughly 750,000 words, and for comparison the entire Lord of the Rings trilogy runs about 500,000 words. A large codebase, say 300,000 lines of Python with comments, fits comfortably inside a single Claude Opus 4.6 context. You could feed in an entire company’s engineering documentation, a year of customer support tickets, or a complete legal case file and ask questions about the whole thing at once.

Previous Claude models maxed out at 200K tokens for input. That’s still large by most standards, but it meant chunking big documents and losing the coherent whole. Opus 4.6 removes that limit for most real work.

The 128K output ceiling matters just as much. Most current models cap outputs at 4K to 8K tokens, and a few hit 32K. With 128K you can ask Opus 4.6 to write a complete document, generate a full module of code, or produce a long analysis in a single response.

SWE-bench: 80.8%

SWE-bench is currently the most credible measure of a model’s software engineering ability. It presents real GitHub issues from popular open-source repositories and asks the model to generate a code fix that passes the associated test suite. Hitting 80.8% on SWE-bench Verified is a real jump. For context, the original GPT-4 resolved just 1.7% of issues when SWE-bench first launched in 2023, and OpenAI’s o3 had been the previous leader at around 71%.

Opus 4.6 at 80.8% means it fixes four out of five real software engineering problems it’s given. This isn’t “write a for loop” territory. These are actual bugs in production codebases that required reading multiple files, understanding context, and producing correct patches.

For developers using AI coding tools, this number matters because it directly translates to autonomous agent performance. An 80.8% pass rate means you can set up a Claude Opus 4.6 agent to work through a backlog of well-defined issues and expect most of them to come back done correctly.

What Changes for Developers

The combination of 1M context and high SWE-bench performance makes a few workflows practical that weren’t before, and here are the ones I’ve been thinking about most.

Repository-scale refactoring. Load an entire codebase into context, describe a refactoring goal, and let Opus 4.6 reason about the full dependency graph before producing a plan. No more chunked analysis losing sight of cross-file relationships.

Long-session agents. Agentic coding tools that run for a while pile up context: file reads, tool outputs, error messages. With 1M tokens, a Claude agent can run much longer before it hits the limit and loses track of what it’s done.

Document analysis at scale. Legal, financial, and research workflows that involve hundreds of pages can now be handled in a single pass. Compare annual reports, analyze full clinical trial data, review entire contract sets in one query.

Pricing

Opus 4.6 sits at the top of Anthropic’s pricing tier, as expected. At launch it was $5 per million input tokens and $25 per million output tokens. That gets expensive at high volume, but for tasks where quality matters more than cost, it’s the right tool, and plenty of tasks fit that bill.

Claude Sonnet 4.6, released shortly after, offers a faster and cheaper alternative for workflows that don’t need the full Opus capability ceiling.

How It Stacks Up Against the Competition

At release, Opus 4.6’s 80.8% SWE-bench score sat clearly ahead of the published scores for earlier OpenAI, Gemini, and DeepSeek models. It would be tested against Google’s Gemini 3.1 Pro, which released two weeks later. For the moment, it held the SWE-bench record.

At launch, the 1M context window also exceeded what OpenAI’s production models offered. Google’s Gemini 1.5 Pro previously held the practical record at 1M tokens, and Gemini 3.1 Pro matched it. The context race has reached a point where the question is shifting from “how many tokens can it fit?” to “how well does it actually use them?”

Anthropic has spent several versions improving long-context retrieval accuracy: the model’s ability to find and use information buried deep in a large context instead of defaulting to what appeared near the beginning or end. Early testing suggests Opus 4.6 handles this well, though I’d expect more systematic evaluations to follow.