Anthropic’s Claude Opus 5 competes on cost rather than raw capability

TL;DR:

• Anthropic has released Claude Opus 5, priced identically to its predecessor at $5 and $25 per million input and output tokens.
• The company positions it near frontier performance at roughly half the cost of its strongest model, Fable 5.
• Anthropic says it is the most aligned model it has tested, and deliberately behind Mythos 5 on cyber capability.

Anthropic has released Claude Opus 5, and the framing is unusually restrained for a frontier launch: the pitch is efficiency, not a capability jump. Pricing holds at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8, while the company claims substantially improved performance at that same price.

Cost per task, not benchmark position

The claims that matter to buyers are comparative cost. On the CursorBench 3.2 coding evaluation, Anthropic says Opus 5 performs within 0.5% of Fable 5’s peak score at half the cost per task. On the OSWorld 2.0 computer-use benchmark it reports surpassing Fable 5’s best result at just over a third of the cost. Customer testimonials emphasise token economy rather than ceiling — one legal-workflow partner reports comparable results with 26% fewer tokens; a trading firm reports roughly a seventh of the reasoning tokens.

On safety, Anthropic says an automated behavioural audit found Opus 5 its most aligned model to date, with the lowest rates of deceptive behaviour. It states the model does not advance the frontier in dual-use capability, remaining behind Mythos 5 in biology and offensive cyber. The distinction it draws is precise: Opus 5 approaches Mythos 5 at finding vulnerabilities but stays substantially behind at exploiting them. Cyber classifiers block binary-based scanning, penetration testing and exploit generation.

Looking forward

For UK businesses, the interesting signal is commercial. When a lab holds price flat and sells efficiency, the competitive pressure has shifted from what a model can do to what it costs to run at volume — the same argument Microsoft and Palantir have made for open weights, and one that two dozen firms pressed on lawmakers this week. That should help UK firms where AI adoption has stalled on unit economics rather than capability. The vendor benchmarks remain vendor benchmarks.