Anthropic’s most powerful model can find tens of thousands of software vulnerabilities and write working exploits for them. So you can’t have it. Today’s release of Claude Opus 4.7 is explicitly positioned beneath that ceiling.
That framing is unusual. Most model launches come with superlatives. This one comes with an asterisk: Opus 4.7 is better than Opus 4.6, better than most competing models on benchmarks, and deliberately less capable than Claude Mythos Preview, which Anthropic has been sitting on since at least early April over cybersecurity concerns. Cyber capabilities in Opus 4.7 were intentionally reduced to make a broad release viable.
What’s actually in the box is worth taking seriously. On a 93-task coding benchmark, Opus 4.7 resolves 13% more problems than Opus 4.6, including four classes of tasks the earlier model couldn’t handle at all. Vision processing jumps to 2,576 pixels on the long edge — more than three times the previous ceiling. On document reasoning, the model makes 21% fewer errors than its predecessor. These aren’t incremental tweaks.
For developers running extended automated workflows, two new features matter. A new xhigh effort level gives finer control over the reasoning-versus-latency tradeoff, useful for complex tasks where waiting is acceptable. Task budgets, now in public beta, let you specify how many tokens Claude can spend across a long multi-session job, which helps prevent runaway costs in production. The /ultrareview command in Claude Code adds a dedicated path for deeper code review sessions.
Pricing is unchanged: $5 per million input tokens, $25 per million output tokens, across Claude.ai, the API, Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry.
There’s one line in the release notes worth flagging. An updated tokenizer means the same inputs now require 1 to 1.35 times more tokens than with Opus 4.6. The per-token price didn’t change. The effective cost for large workloads did.
Anthropic says releasing Opus 4.7 builds the foundation for eventually distributing Mythos-class capabilities safely. That’s the goal. The actual schedule for when Mythos reaches general availability remains unstated.
This article is AI-generated.
