Claude Opus 5 makes proactivity the central agent metric

Anthropic has introduced Claude Opus 5 for everyday coding and knowledge work. The company says the model approaches Claude Fable 5 capability at half the price and more than doubles Opus 4.8 performance on Frontier-Bench v0.1 at a lower cost per task. Simon Willison highlighted one phrase in particular: thoughtful and proactive.

Opus 5 built the missing tool inside a benchmark task

Anthropic describes a task in which the model received an image of a machine part without a direct tool for viewing it. The company says Opus 5 wrote a postup počítačového vidění, extracted geometry from pixels and reconstructed the component in FreeCAD.

That example tests something different from answering a well-prepared prompt. The model identified an obstacle, created an intermediate step and continued toward the result. This behavior is valuable in long agent tasks where every dependency cannot be prepared in advance.

Companies are buying completed work, not reasoning steps

Reported pricing is $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. Fast mode costs twice the base rate. A team should therefore care about cost per completed task, the number of human interventions and the share of proposals that pass tests.

Proactivity can reduce delays caused by requests for more instructions. It also creates more model decisions that need to appear in traces and remain legible during review. The saving is real only when autonomy does not generate a larger verification burden at the end.

Model initiative amplifies ambiguity in the assignment

Anthropic calls Opus 5 its most aligned model and says it withheld some training related to high-risk cyber capabilities. Stronger alignment still cannot resolve an ordinary failure mode: an agent may execute a badly framed request with great competence. Production systems continue to need tool limits, tests and approval points.

Review queues and reverted changes will provide the useful measure

A strong signal would be migrations, fixes and analyses completed with fewer human interventions and no rise in incidents. If the only change is a larger diff and a longer review, proactivity becomes an expensive way to produce more work for the reviewer.

Lilith's verdict

Opus 5 will reveal itself when a reviewer opens a large diff the model produced without asking follow-up questions. Its value is the number of changes that person can actually approve after inspection.