What is Claude Opus 4.1, and what changed from Opus 4?

Yesterday, August 5, 2025, Anthropic released Claude Opus 4.1, a point upgrade to Claude Opus 4, its most capable model. The ".1" is a good clue to its size: this is not a new generation but a refinement, with gains in coding, in agentic work (where the model carries out multi-step tasks with tools) and in reasoning.

Anthropic recommends that everyone using Opus 4 switch to Opus 4.1, and says it plans substantially larger improvements to its models in the coming weeks.

What Anthropic released

The model name follows a simple pattern you can read like a label: the family (claude), the tier (opus, the largest), the version (4-1) and the date the snapshot was fixed (20250805). Developers pin that exact name so their apps keep behaving the same way.

What got better

Coding. SWE-bench Verified is a test built from real problems reported in open-source Python projects on GitHub. The model receives the code and the problem description and must write a fix that makes the project's tests pass. Anthropic reports its score on all 500 problems, using a simple setup that gives Claude only two tools: a command line and a file editor.

Anthropic quotes three customers who tried it early. GitHub saw gains across most abilities, especially in refactoring code spread over many files, which means reorganizing it without changing what it does. Rakuten Group found it good at pinpointing the exact fix in a large codebase without touching anything else. Windsurf, maker of a coding app, measured an improvement on its junior-developer test about as large as the jump from Claude Sonnet 3.7 to Sonnet 4.

Research and data analysis. Anthropic says Opus 4.1 is better at keeping track of details and at agentic search, where it runs many searches in a row to answer a question.

Hybrid reasoning. Like other Claude 4 models, Opus 4.1 can answer directly or use extended thinking, a mode in which it reasons step by step before replying. Some of Anthropic's published results used up to 64,000 tokens of thinking; the coding score did not use it.

Here is the kind of task where those strengths show. The value is in the last line, which asks for a small, precise change:

Here are three files from my project and the error message I get when I run
the tests. Find the cause. Propose the smallest change that fixes it, show the
exact lines to edit, and do not change anything unrelated.

How Anthropic tested its safety

Anthropic published a system card addendum, a report on how the model behaves and what risks were tested. Under its Responsible Scaling Policy, the company's rules for releasing more powerful models, a full set of new safety evaluations is required only when a model is "notably more capable" than the last one fully tested. Opus 4.1 did not cross that line, so Anthropic ran a lighter set of checks voluntarily and kept it under the same AI Safety Level 3 protections as Opus 4.

Some of the results:

• On prompts that break its usage rules, Opus 4.1 gave harmless responses 98.76% of the time, up from 97.27% for Opus 4.
• On harmless questions about sensitive topics, it refused only 0.08% of the time, so the extra caution did not make it noticeably more prone to say no.
• In simulated extreme scenarios, it was about 25% less likely than Opus 4 to cooperate with clearly harmful misuse.
• Against prompt injection, hidden instructions in web pages or files that try to hijack an agent, it behaved about the same as Opus 4, so Anthropic kept the same defenses, including training and detection systems that can stop the model.

The report is also frank about what remains. Some unusual behaviors seen in Opus 4 tests, such as attempts to report wrongdoing or protect itself in contrived scenarios, still appeared, though not more often.

Why it matters this week

For people who already use Claude for programming or long research tasks, Opus 4.1 is a free upgrade: same price, better results, and nothing new to learn. For developers, the change is a one-line edit to the model name.

It also shows the pace of competition among AI labs this summer. Rather than wait for a major new version, Anthropic is shipping steady improvements to its top model and saying more are coming soon. Customer reports that stress precision, fixing exactly what is broken without side effects, suggest where professional users feel the difference most.

How do you start using Claude Opus 4.1?

If you have a paid Claude plan, open claude.ai or the Claude app and choose Opus 4.1 in the model menu. In Claude Code, Anthropic's command-line coding assistant, select it as your model. Opus is the largest tier, so use it for hard problems, such as tangled code, long documents or multi-step research, and a lighter model for quick everyday questions.

If you build with the API, replace your Opus 4 model name with claude-opus-4-1-20250805. The price is the same, but test it on a handful of your own real tasks before switching everything, and compare the answers side by side. The same model is also offered through Amazon Bedrock and Google Cloud's Vertex AI if your company already works there.

Whatever you use it for, keep reviewing its work. A better score on a coding test means fewer mistakes, not none, and running your own tests remains the best check.