Anthropic unveiled Claude Opus 4.7 on 16 April 2026, calling it its strongest general availability model for software engineering, enterprise workflows, and long-running agent tasks. It is not the strongest model in the house, because that role remains with Mythos Preview, but it is the one the company has decided to really put in the hands of the market, with availability on Claude, API, Amazon Bedrock, Google Vertex AI and Microsoft Foundry at the same prices as Opus 4.6: $5 per million tokens in input and $25 per million in output. The move is significant because it moves the boundary between coding assistant and semi-autonomous technical colleague even further.
The central point is not just that Opus 4.7 writes better code. Anthropic insists above all on three qualities: it follows instructions more precisely, handles complex and lengthy tasks more consistently, and checks its output better before returning it. In practice, it means less need for human babysitting on articulated debugging, deep reviews, legacy migrations, and jobs distributed over multiple files. The first testers cited by Anthropic speak openly of a leap in reliability, with fewer errors on intermediate steps, better tool management and a more pronounced tendency to correct itself when it intercepts a logical flaw.
Then there is a second, less noisy but very important front: vision. Opus 4.7 can process images up to 2,576 pixels on the long side, more than three times the capacity of previous Claude models, improving use on dense screenshots, technical diagrams, complex interfaces and multimodal tasks where visual detail makes a difference. Anthropic also claims that the model has a more mature visual flavour in professional work, such as slides, documents and interfaces. This is an interesting clue: coding is no longer separate from design, but enters a hybrid zone where understanding a layout and turning it into a product matters almost as much as writing a function.
The best practices published for Claude Code tell a good story about the kind of model Opus 4.7 wants to be. Anthropic recommends treating it more like an engineer to be delegated a complete task than a pair programmer to be guided line by line. The starkest suggestion is to put relevant targets, constraints, acceptance criteria and files already in the first prompt, because each additional round introduces reasoning overhead and thus more token consumption. Translated: less micromanagement, more well-constructed initial brief.
The effort control also changes. The new default level in Claude Code is ‘xhigh‘, intermediate between high and max, and is indicated by Anthropic as the best compromise for most agentical coding tasks. The max level remains available for really hard problems, but with diminishing returns and a greater tendency to ‘overthink’, while high is recommended when costs need to be kept down or several sessions need to be run in parallel. In addition, Opus 4.7 no longer supports fixed-budget Extended Thinking: in its place comes adaptive thinking, which dynamically decides when to investigate and when to respond directly.
Then there are other operational changes that teams should not underestimate. Opus 4.7 tends to use fewer tools, generate fewer sub-agents by default and be less verbose on simple requests, calibrating the length according to the complexity of the task. This is why Anthropic warns that prompts optimised for Opus 4.6 may give unexpected results: the new template takes instructions more literally and must be retuned. This is the classic sign of models that are really starting to mature: you no longer just have to learn what they can do, you have to learn how to delegate properly.
In parallel, Anthropic has also added new elements of control and security: task budgets in public beta on the API, the /ultrareview command for dedicated code reviews, the extension of auto mode to Max users, and an automatic safeguards system that blocks high-risk cyber requests, while legitimate security professionals can log in via the Cyber Verification Program. This is not just a benchmark upgrade. It is an attempt to put a more autonomous model into production without turning autonomy into a problem. And that, today, is much more difficult than scoring a few extra points on a leaderboard.
Note: https://www.anthropic.com/news/claude-opus-4-7
Note best practices: https://claude.com/blog/best-practices-for-using-claude-opus-4-7-with-claude-code
