Claude Opus 4.7 is here
Anthropic just dropped Opus 4.7. Same price as 4.6 ($5 input, $25 output), but noticeably smarter on the long-running stuff agents actually do.
What's different. A new xhigh effort level lets you dial in reasoning depth when the task is gnarly. Task budgets, now in public beta, cap token spend across multi-step runs — the thing every team building on Claude has been hacking together by hand. Vision jumps to 2,576 pixels on the long edge, roughly 3.75 megapixels. That's about 3x the old ceiling, which matters for document-heavy agents.
Anthropic also quietly shipped /ultrareview, a slash command for dedicated code review sessions inside Claude Code. Feels like a trial balloon for specialized modes inside Code.
The benchmarks Anthropic is flexing: state-of-the-art on GDPval-AA (their economically valuable knowledge work eval) and highest-ever on the Finance Agent evaluation. Two things that matter for where this goes — knowledge work automation and finance agents. Not coincidentally, the two verticals where AI spending is most real right now.
One note: Anthropic explicitly reduced cybersecurity capabilities compared to the internal Mythos Preview. They're running a Cyber Verification Program for legit security pros to get unlocked access. Interesting direction — not pretending the model lacks offensive capacity, they're gating it.
Model ID: claude-opus-4-7. Available on Anthropic API, Bedrock, Vertex, Foundry. https://www.anthropic.com/news/claude-opus-4-7
← Back to all articles
What's different. A new xhigh effort level lets you dial in reasoning depth when the task is gnarly. Task budgets, now in public beta, cap token spend across multi-step runs — the thing every team building on Claude has been hacking together by hand. Vision jumps to 2,576 pixels on the long edge, roughly 3.75 megapixels. That's about 3x the old ceiling, which matters for document-heavy agents.
Anthropic also quietly shipped /ultrareview, a slash command for dedicated code review sessions inside Claude Code. Feels like a trial balloon for specialized modes inside Code.
The benchmarks Anthropic is flexing: state-of-the-art on GDPval-AA (their economically valuable knowledge work eval) and highest-ever on the Finance Agent evaluation. Two things that matter for where this goes — knowledge work automation and finance agents. Not coincidentally, the two verticals where AI spending is most real right now.
One note: Anthropic explicitly reduced cybersecurity capabilities compared to the internal Mythos Preview. They're running a Cyber Verification Program for legit security pros to get unlocked access. Interesting direction — not pretending the model lacks offensive capacity, they're gating it.
Model ID: claude-opus-4-7. Available on Anthropic API, Bedrock, Vertex, Foundry. https://www.anthropic.com/news/claude-opus-4-7
Comments