TL;DR:

• Anthropic has released Claude Opus 4.7 to general availability with a 13% lift on its internal coding benchmark and new automated blocks on prohibited cyber requests.
• The company is using Opus 4.7 as a live testbed for safeguards it ultimately wants to deploy on the restricted Mythos-class models.
• Pricing holds at $5 per million input tokens and $25 per million output tokens, same as Opus 4.6.

Anthropic has made Claude Opus 4.7 generally available across its own products, the Claude API and the major clouds — Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry. The release comes a week after the company unveiled Project Glasswing, its framework for handling the cyber risk exposed by the more capable Mythos Preview.

A deliberately less-capable sibling

Opus 4.7 is framed as the model where Anthropic tests the guardrails it eventually wants to ship on Mythos. During training, the company says it ran experiments to differentially reduce cyber capability. The released version ships with automated systems that detect and block prompts signalling prohibited or high-risk cybersecurity use, while legitimate researchers can apply for the new Cyber Verification Program to unlock red-teaming and vulnerability-research workflows.

The capability gains are concentrated in long-running software engineering. Anthropic reports Opus 4.7 resolved 13% more tasks than Opus 4.6 on its 93-task internal coding benchmark and solved four tasks neither Opus 4.6 nor Sonnet 4.6 could handle. Early-access feedback from fintech and data-tooling partners points to stronger self-verification, better instruction following and — unusually — a more opinionated model that pushes back on weak user premises rather than agreeing with them.

UK developer angle

For UK teams, Opus 4.7 lands at the same moment as Sovereign AI’s push to put British AI builders on equal footing with US labs. Pricing is unchanged, which makes the upgrade a simple drop-in for teams already running Opus 4.6. The cyber-misuse blocks are more consequential for British security firms: tooling that previously needed legal opinions to justify access now has a formal verification route, which should help UK penetration-testing and threat-intelligence companies integrate Claude without reputational risk.

Looking forward

Anthropic’s staggered pattern — release a guardrailed sibling, watch how the safeguards perform in the wild, then widen access to Mythos-class capability — is the company’s working answer to the regulatory pressure that Mythos has surfaced in London, Washington and Brussels. If Opus 4.7’s automated blocks hold up, Mythos is likely to follow. If they do not, expect a tighter release policy from every frontier lab that was watching.