The case Anthropic makes on benchmarks
The pitch centers on sustained, self-checking work rather than single-turn chat quality: Anthropic says Opus 5 is better at verifying its own outputs, iterating after a failure and carrying multi-step tasks to completion. On Anthropic's own software-engineering numbers, Opus 5 leads Frontier-Bench v0.1 and more than doubles Opus 4.8's score at a lower operational load per task, and lands within 0.5% of Fable 5's peak on CursorBench 3.2 at maximum effort while costing half as much per task.
The knowledge-work figures follow the same shape. Anthropic reports Opus 5 scoring three times as high as the next-best model on ARC-AGI 3, passing about 1.5 times as many tasks at equal operational load on Zapier AutomationBench, and beating every other model at any given operational load on OSWorld 2.0. In life sciences it says Opus 5 improves on Opus 4.8 across structural biology, organic chemistry and bioinformatics, with gains of 10.2 percentage points on an internal organic chemistry benchmark and 7.7 percentage points on a protein-related task.
Every one of these is a vendor number. Anthropic's own footnote describes the Frontier-Bench results as an internal run on the mini-SWE-agent harness with a GKE backend, averaging reward over five attempts per task and falling back to Opus 4.8 when the safety classifier refused an Opus 5 or Fable 5 request. Long-running agents tend to fail in ways benchmarks with five retries do not surface well: losing context, stopping early, skipping verification, or committing to a wrong first approach and never recovering. Those are exactly the behaviors Anthropic says it improved, which is why the useful evidence will come from teams running Opus 5 against their own multi-file changes, repository-aware code review and tool-using research before moving production workflows off Opus 4.8.
Where the guardrails draw the line
Anthropic says Opus 5 does not advance the frontier in dual-use risk. It remains behind Mythos 5 on offensive cybersecurity and biology research, even though broader capability gains pushed its cyber performance up substantially. On finding vulnerabilities Opus 5 comes close to Mythos 5; on developing working exploits it stays well behind. The deployed classifiers encode that boundary directly: they permit vulnerability finding in source code but block binary-based vulnerability scanning, penetration testing and exploit generation.
Flagged requests in Claude.ai, Claude Code and Claude Cowork fall back to Opus 4.8 by default, and API callers can enable the same fallback. On biology, Anthropic calls Opus 5 its most capable generally available model for scientific research while stating it still has important limitations on long-running autonomous research, the setting where it expects the most serious biological risk to concentrate and where Mythos 5 remains stronger. The open question for developers is whether that routing quietly reshapes behavior on legitimate security and research work, since a request that silently drops to Opus 4.8 will not perform like the model they think they are calling.