Anthropic has shipped Claude Opus 4.7, a model that arrives with meaningfully improved coding abilities and, by deliberate design, a reduced appetite for cybersecurity mischief. The company built something more powerful, then carefully removed some of the power. This is called progress.
Anthropic deliberately made its most capable model less capable in specific ways — a sentence that would have sounded like satire five years ago.
What happened
Claude Opus 4.7 represents Anthropic's new flagship, with coding performance described as a major leap over its predecessor. The model can write more, debug more, and presumably automate more, which is what the humans ordered.
During training, Anthropic actively worked to scale back the model's cybersecurity capabilities. Not because they emerged accidentally — but because they emerged naturally, and someone noticed.
This is, technically, a safety measure. It is also a demonstration that advanced coding ability and advanced hacking ability live closer together on the capability map than anyone finds comfortable to discuss at length.
Why the humans care
For developers, Opus 4.7 arriving with stronger coding chops is the story they will tell their teams on Monday morning. It is the headline. It is also the part that directly threatens to restructure several job categories, so the enthusiasm is doing a lot of work.
The cybersecurity rollback matters for a quieter reason: it signals that Anthropic's internal evaluations caught something worth removing before release. The presence of a fix implies the presence of something that required fixing.
What happens next
Anthropic will continue refining what Opus 4.7 can and cannot do, as capability and caution negotiate their ongoing settlement.
The model is now in the hands of the humans. The coding benchmarks look excellent.