AI company Anthropic presents Claude Opus 4.7 as a model that can carry more of a task through to completion, instead of only starting it. The company says it also checks its own work before reporting back. Businesses could use it from April 2026, while Mythos Preview, the company’s model that drew so much attention online, was limited to a programme called Project Glasswing.
The main gain is reliability over longer stretches of work, and Anthropic’s early testers point the same way. Testers at Devin say the model can work coherently for hours and keep pushing through difficult problems. At Notion, it keeps going when a software tool it relies on fails, a problem that used to stop earlier versions. At Factory, it carries tasks through instead of stopping halfway.
Coding sits at the centre of that case. Opus 4.7 improves on the earlier Opus 4.6 in advanced software engineering, especially on the hardest tasks. The test that best shows this uses software tasks that can take engineers hours or days and often require changes across several files. Work like that creates delay, rework and management overhead inside real teams, so it makes a more useful test than a short demo.
Precision is the second change. Anthropic says Opus 4.7 follows instructions more to the letter than Opus 4.6. That is a strength, but instructions written for older models may need revising. Loose ones are more likely to fail visibly, while tighter ones with clearer context are more likely to hold up.
Vision and memory improve too. The model supports images at more than triple the previous resolution, which the company ties to screenshots, complex diagrams and tasks where the model operates a computer. It is also better at keeping memory in files, such as saved notes and task files it can reuse across sessions. More of the material teams already work with can be used directly, and more context carries forward without being restated every time.
Finance follows the same pattern. Anthropic says the model beats Opus 4.6 on financial analysis, with more rigorous analyses, stronger financial models and more professional presentations. The test behind that claim uses financial research tasks built around SEC filings and questions written by experts. Those tasks sit close to real analyst work, so the improvement shows in what the model hands back, in a field where mistakes are expensive.
