کلوڈ اوپس 5

Anthropic's model for complex agentic coding and enterprise work, described by the lab as a step change over Opus 4.8 rather than an incremental one, reaching close to Fable 5's frontier intelligence at half the price. Its largest gains are in deep reasoning, long-horizon agentic tasks and test-time compute scaling, with strong results in code review and bug finding, vision, long-context work and document tasks. Thinking is adaptive and on by default, with effort selectable from low up to max.

Anthropic released Opus 5 as the model to reach for every day rather than the one held back for the hardest hour: it became the default on Claude Max and the strongest model offered on Claude Pro on the day it shipped. The lab positions it as coming close to Claude Fable 5's frontier intelligence at half the price, and it charges exactly what Opus 4.8 charged before it, so the generational gain arrives without a rate change.

Anthropic presents almost all of its results as cost-per-task curves across the effort dial rather than as single peak scores, which is the honest way to read a model whose thinking budget the caller sets. On Frontier-Bench v0.1, its software engineering evaluation, the lab reports Opus 5 ahead of every other model and more than double Opus 4.8 at a lower cost per task. On CursorBench 3.2 at max effort it lands within 0.5% of Fable 5's peak score for half the cost per task, and beats every other model at a given cost on the high, xhigh and max rungs.

The same pattern holds away from code. Anthropic reports three times the next-best model's score on ARC-AGI 3, which scores models on problems they have not seen before; a pass rate around 1.5 times the next-best model on Zapier's AutomationBench at the same cost per task, with more tasks passed than any other model even at the lowest effort setting; and better results than any model at any cost on the OSWorld 2.0 computer-use evaluation, passing Fable 5's best score at just over a third of the cost. It is also the lab's best and most cost-efficient model on GDPval-AA v2, HLE AutomationBench and DeepSearchQA.

What early-access testers described was less a score than a working habit: the model checks itself. Anthropic's own examples include a Frontier-Bench task where the model was given a mechanical drawing and deliberately no way to view it, and answered by writing a computer vision pipeline to recover the geometry from raw pixels; a real bug in a widely used package manager where it found the root cause an accepted community patch had missed; and a market data feed built in one session, with its own test harness written because no live feed existed to validate against. On science the lab reports Opus 5 ahead of Opus 4.8 on every one of its internal life-sciences evaluations, by 10.2 points on inferring molecular structures from spectroscopy and 7.7 points on predicting how protein sequence variants behave.

Anthropic is also specific about where the model stops. It says Opus 5 does not advance the frontier in dual-use capability: on its OSS-Fuzz evaluation it comes close to Mythos 5 at finding software vulnerabilities but stays far behind at turning them into working exploits, and Mythos 5 remains the stronger model for long-running autonomous biology research. Its safety classifiers allow source-code vulnerability review while blocking binary-based scanning, penetration testing and exploit generation. On the lab's automated behavioural audit it scored 2.3 for overall misaligned behaviour, the lowest of any recent Anthropic model.