Opus 5 wrote its own computer vision code

Opus 5 wrote its own computer vision code to reconstruct a 3D model from a machine part drawing it couldn't directly see, per Simon Willison's July 24 review.

That's the detail worth sitting with if you send Opus 5 messy or incomplete inputs. It didn't ask for better data. It wrote code to pull geometry out of raw pixels and modeled the part itself.

Willison is unusually hedged about the rest. He says he was kayaking when Anthropic shipped the model and hadn't run his usual tests yet. He does flag that the faster variant costs twice as much as the standard one.

He also passes along Anthropic's framing on security. Opus 5 got better at finding vulnerabilities through general capability gains, not because it was trained on exploitation tactics. Willison ties the distinction to the ExploitGym incident without explaining what happened there, so the caveat reads more like reassurance than evidence.

The take

The CV demo is the flashy part, but the vulnerability-finding line is the one worth watching. A model that finds more exploits every generation without being taught exploitation is a promise, not a guarantee. That promise gets harder to keep as the capability gap widens. Anthropic's own post is the only source saying it's held so far.