Anthropic releases Claude Opus 4.7 for agentic coding, vision and autonomous tasks

On April 16, 2026, Anthropic released Claude Opus 4.7, a targeted but very noticeable update to its flagship model. It aims not at abstract benchmarks but at developers' practical pain points: agentic coding, work with high-detail images, and long autonomous tasks where the model must not just answer but plan, check itself and see the process through to the end. The company positions Opus 4.7 as a direct upgrade to Opus 4.6, not a new model line. In engineering scenarios, though, the difference turned out to be substantial.

According to Anthropic, the model follows instructions better, handles multi-step workflows more carefully, and more often finds a way to verify its result before returning it to the user. This matters most in settings where AI does not write a single snippet of code on request but acts as a semi-autonomous worker: it reads a repository, runs tools, fixes errors, keeps going after failures and holds intermediate decisions in memory.

On benchmarks from Anthropic and its partners, the gains do not look cosmetic. On an internal 93-task coding benchmark, Opus 4.7 improved the share of solved tasks by 13% over Opus 4.6, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve. On CursorBench, the model scored 70% versus 58% for the previous version. In Notion's cases for complex multi-step workflows, the gain was 14%, with fewer tokens used and a third as many tool-calling errors. Rakuten separately claims that on Rakuten-SWE-Bench the new model resolves three times as many production tasks as Opus 4.6.

The second major upgrade is vision. Claude Opus 4.7 can accept images up to 2576 pixels on the long side, roughly 3.75 megapixels. That is more than three times the limit of previous Claude models. In an ordinary chat mode the figure seems secondary, but for agentic scenarios it is critical: the model can parse dense interface screenshots, technical diagrams, schematics, chemical structures and documents where small elements matter. Anthropic cites a computer-use example: on the XBOW visual benchmark, the new version scored 98.5% versus 54.5% for Opus 4.6. In practice, this moves a number of tasks from the "works sometimes" category to the "can be built into a product" category.

Anthropic has also strengthened tools for long autonomous execution. Opus 4.7 adds a new reasoning effort level, xhigh, sitting between high and max. In the API, the company launched a public beta of task budgets, so developers can cap the token budget on long runs and manage step priority. Claude Code gains an ultrareview mode for deeper review of changes, and auto mode now interrupts long tasks with permission requests less often.

Migrating from Opus 4.6 is not entirely free, however: the updated tokenizer can increase input token counts by roughly 1.0 to 1.35 times depending on content type, and deeper reasoning in the later stages of agentic tasks raises output token volume.

Anthropic also stresses safety. Opus 4.7 arrives shortly after the announcement of Mythos Preview, a more powerful model whose access the company chose to restrict because of cybersecurity risks. As a result, Opus 4.7 is the first public model on which Anthropic is testing automatic blocks for high-risk cyber requests. For legitimate uses such as pentesting, red teaming and vulnerability research, the company has opened a verification program in parallel. By Anthropic's own assessment, Opus 4.7's safety profile is broadly close to 4.6: the model improved on honesty and resistance to prompt injection, although it is not perfect in some categories and still trails Mythos Preview in overall behavioral calibration.

The main takeaway is simple: Claude Opus 4.7 is not a new generation for the sake of a new number, but a pragmatic release for those building real workflows on top of the model. If your use case is an IDE assistant, a code review agent, document analysis, computer vision over interfaces, or long chains of actions with tools, the update looks significant right now. For mainstream users the difference may not always be obvious, but for developers and teams that measure quality not by demos but by the number of tasks completed end to end, Opus 4.7 looks like one of Anthropic's most useful updates in a while.