Anthropic has released Claude Opus 4.7, a direct upgrade to Opus 4.6 with improvements in software engineering, vision, and agentic tasks. Available via Claude products, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Pricing stays at $5 per million input tokens and $25 per million output tokens. However, this isn’t quite the much-discussed Mythos model, as Anthropic considers that too risky an KLM to release fully.
Anthropic describes Claude Opus 4.7 as a notable step up from Opus 4.6. It should perform particularly well for advanced software engineering. The company says early users are now handing off their most difficult coding tasks to the model with confidence, something that previously required close supervision. Not only that, it had appeared as a recent chink in Anthropic’s armor.
Mere days ago, developers had been vocal about inconsistencies in Opus 4.6 and Claude Code, with some reporting reliability issues on complex engineering problems. Opus 4.7 is Anthropic’s direct response to that feedback.
However, it isn’t intended to be the fully released equivalent of Mythos, the LLM that Anthropic says is too dangerous to release into the wild. The reason is pretty straightforward if ominous: it can detect vulnerabilities at such a rate, that previously unexploited software weaknesses could emerge from decades of obscurity to bring down the Internet. Or, at any rate, that is what Anthropic and some early testers have suggested.
Sharper vision and better instruction-following
Beyond coding, the actually released Opus 4.7 brings a three-fold improvement in image resolution. It processes images up to 2,576 pixels on the long edge (~3.75 megapixels). This opens up computer-use agents reading dense screenshots, Anthropic says, as well as data extraction from complex diagrams, and pixel-perfect visual references.
It is worth noting that the improved instruction-following can produce unexpected results for prompts written for earlier models. Where previous models interpreted instructions loosely or skipped parts, Opus 4.7 takes them literally. Anthropic recommends re-tuning prompts and harnesses accordingly. Truth be told, this is always a recommendation when switching models.
Cybersecurity safeguards and the Mythos context
Opus 4.7 is the first model to ship with new cybersecurity safeguards coming out of Project Glasswing. Anthropic stated last week that it would keep its Claude Mythos Preview limited due to its unprecedented cyber capabilities, and test new safeguards on less capable models first. Opus 4.7 is that test bed, or at least the first one of such models.
Its cyber capabilities are deliberately kept below Mythos level, and the model automatically detects and blocks requests for prohibited or high-risk cybersecurity uses. The Dutch government had expressed concern about the Mythos model’s risks, and Anthropic’s staged rollout is an acknowledgement of this type of scrutiny. Security professionals can apply for the new Cyber Verification Program to use Opus 4.7 for legitimate purposes such as penetration testing and red-teaming.
New developer tools and migration notes
Alongside the model, Anthropic is shipping several developer updates. A new xhigh effort level sits between high and max, giving finer control over the reasoning versus latency trade-off. In Claude Code, the default effort level has been raised to xhigh for all plans. Task budgets are entering public beta on the API, helping developers manage token spend over longer runs. The new /ultrareview command in Claude Code flags bugs and design issues, Pro and Max users get three free ultrareviews to start.
Developers migrating from Opus 4.6 should account for two changes that affect token usage: an updated tokenizer that can map the same input to roughly 1.0–1.35x wmore tokens, and increased output tokens at higher effort levels on agentic tasks. Anthropic has published a migration guide with further advice.