Claude Opus 4.1 improves AI-powered programming

Anthropic has launched Claude Opus 4.1, an update to its most advanced model for programming, research, and complex tasks. It is now available to paid users of Claude and Claude Code, as well as customers using its API, Amazon Bedrock, and Google Cloud Vertex AI.

The main improvement appears in real-world programming: Claude Opus 4.1 reaches 74.5% on SWE-bench Verified, a test that measures whether a model can solve coding problems drawn from real projects. Anthropic says the model is especially effective when working across multiple files and making precise changes without modifying parts it does not need to touch.

The price remains the same as Claude Opus 4. To use it through the API, developers must select the claude-opus-4-1-20250805 model.

Fewer unnecessary changes when programming

The difference is not only in how Claude writes code. How it modifies an existing project also matters.

GitHub says Opus 4.1 improves on Opus 4 across most capabilities and highlights its results when reorganizing code spread across multiple files. Rakuten Group, meanwhile, says the model identifies specific fixes within large codebases without introducing unnecessary changes or new errors.

In practice, this could help you debug an application without having the AI rewrite functions that already work. That matters when code spans thousands of files and a small change can cause problems elsewhere in the system.

Windsurf, a developer tools platform, reports an improvement of one standard deviation in its test for junior programmers. According to the company, the increase is roughly comparable to the improvement between Sonnet 3.7 and Sonnet 4 in the same environment.

Research also improves

Anthropic says Opus 4.1 advances detailed research and data analysis, especially when tracking specific information and conducting multi-step searches. This type of work is known as an agentic task: the AI does not respond all at once. It decides what to search for, uses tools, and connects results before delivering an answer.

For example, it could help you compare lengthy documents, review a company’s data, or investigate a technical problem without losing important details along the way.

The published results come with some caveats. The 74.5% on SWE-bench Verified was achieved without enabling extended reasoning, while other tests cited by Anthropic, such as GPQA Diamond, AIME, and TAU-bench, used up to 64,000 reasoning tokens. Anthropic and OpenAI also do not report exactly the same number of problems in SWE-bench: Claude is evaluated on all 500 exercises, while OpenAI’s results use a subset of 477.

For most users, the change means a more precise version of Claude Opus 4, not a completely different product. Anthropic recommends upgrading from Opus 4 and says it is preparing larger improvements for the coming weeks. You should check whether those improvements hold up in everyday projects, not only in benchmark tests.