Claude Opus 4.7 improves AI for coding

Anthropic now offers Claude Opus 4.7, an artificial intelligence model focused on solving complex programming and professional tasks with less human supervision. It is available today in Claude, through its API, and on platforms including Amazon Bedrock, Vertex AI, and Microsoft Foundry.

The company presents it as a direct improvement over Opus 4.6, especially for long tasks that require several steps: analyzing code, finding an error, proposing a change, testing it, and verifying that the solution works. The idea is for the model to do more than respond. It should be able to keep working for longer without losing context or stopping at the first problem.

More autonomy for programming tasks

In an internal test of 93 coding tasks, Opus 4.7 solved 13% more tasks than Opus 4.6. It also completed four problems that neither its predecessor nor Sonnet 4.6 had been able to solve, according to Anthropic.

Early users describe improvements in several areas:

• Fewer errors when using tools.
• A greater ability to review its own work.
• Better adherence to precise instructions.
• More consistency on tasks that run for a long time.
• Better recovery when a tool fails.

On CursorBench, a programming test, Opus 4.7 scored more than 70%, compared with 58% for Opus 4.6. In another real-world task environment, Rakuten says it solved three times as many production issues as the previous version. These figures come from evaluations by Anthropic and partner companies, so they are not equivalent to a single independent test applied to all models.

For you, the practical change is simple: tasks that once required checking every step can move to a more autonomous workflow. For example, you can ask it to investigate a bug in an application, modify several files, run tests, and give you an explanation of what changed. The generated code still needs review, especially in critical systems.

It also understands images and documents better

Opus 4.7 improves its ability to interpret high-resolution images. It can accept images of up to 2,576 pixels on the longest side, or about 3.75 megapixels, more than three times the resolution supported by earlier Claude models.

This is useful for reading dense screenshots, technical diagrams, chemical structures, or documents where small details matter. It can also help create interfaces, presentations, and documents with a more polished finish, according to Anthropic's and its partners' initial tests.

The improvement comes at a cost: more detailed images use more tokens, the units the model uses to process text and images. If you do not need visual precision, reducing the image size can lower consumption.

Same price, but more control over effort

The price remains 5 dollars per million input tokens and 25 dollars per million output tokens, the same as Opus 4.6. Developers can use the model through claude-opus-4-7.

Anthropic is adding an effort level called xhigh, positioned between high and max. It lets you decide how much time and compute the model dedicates to reasoning: a higher level usually improves results on difficult problems, but increases latency and consumption.

In Claude Code, Anthropic's programming tool, xhigh becomes the default level. Anthropic is also introducing /ultrareview, a command that creates a dedicated session to review changes and look for code errors or design problems.

More capability does not mean total freedom in cybersecurity

Anthropic clarifies that Opus 4.7 is less capable in cybersecurity than its Claude Mythos Preview model. The company is using this version to test its new safeguards first, before expanding access to more powerful models.

The model includes systems that automatically detect and block requests related to prohibited or high-risk uses. Professionals working on vulnerability research, penetration testing, or security exercises can request access to the new Cyber Verification Program.

Anthropic's safety evaluation describes Opus 4.7 as generally reliable, though not perfect. It improves in honesty and resistance to certain instruction attacks, but performs slightly worse in some cases, such as its tendency to give overly detailed advice about controlled substances.

What current users should keep in mind

Opus 4.7 is compatible with Opus 4.6 workflows, but Anthropic recommends reviewing your configurations. The model follows instructions more literally, so an ambiguous instruction that was previously interpreted flexibly may now produce an unexpected result.

In addition, the new tokenizer can convert the same text into between 1 and 1.35 times more tokens, depending on the content. In long tasks, higher effort levels can also generate more text. Developers will be able to control that consumption through the effort level, task budgets, and more concise prompts.

Anthropic's move points to a clear trend: models are not just answering questions. They are also executing complete processes with tools, memory, and intermediate reviews. Opus 4.7 aims to let you delegate longer technical tasks, but the decisive question will be whether that autonomy remains stable once it leaves controlled tests and faces everyday work.