Anthropic released Claude Opus 5. It costs half of Fable 5 and outperforms it in some tests

In the hierarchy of language models, a simple rule applies: the more expensive, the smarter. Anthropic somewhat broke this rule with today’s release of Claude Opus 5. The new model costs 5 dollars per million input tokens and 25 dollars per million output tokens (approximately 105 and 525 CZK), exactly half of what the company charges for the flagship Fable 5. And in some of its own tests, it beats the more expensive model.

Cheaper sibling that outgrew the flagship

Anthropic introduces Opus 5 as a new standard for programming and knowledge work. On the Frontier-Bench v0.1 benchmark, according to the company, it more than doubles the performance of the previous Opus 4.8, at a lower cost per task. On the computer control test OSWorld 2.0, it surpasses the best result of Fable 5 at roughly one-third the cost. And on ARC-AGI 3, which measures the solution of previously unseen problems, it is said to be three times more successful than the second model in ranking.

Here’s the obligatory reminder: all these numbers come from charts compiled by Anthropic itself. Independent verification is not yet available, and for benchmarks in general, each manufacturer chooses those in which they look good. However, the price/performance ratio is so skewed that a rapid shift of users towards the cheaper model can be expected.

Opus 5 also becomes the default model for the Claude Max subscription and the most powerful available in the Pro plan. It offers a context window of 1 million tokens, a maximum output of 128 thousand tokens, and a knowledge cutoff of May 2026. A Fast mode is also available, which runs approximately 2.5× faster at double the price.

Model that wrote its own tool

More interesting than tables are the specific examples Anthropic boasts about. In one test task, the model was given a drawing of a machine part with the task of rewriting it into a 3D model in FreeCAD – but intentionally without the ability to view the drawing. Opus 5 therefore wrote its own computer vision pipeline, extracted the geometry directly from pixels, and reconstructed the part. According to the company, no competing model repeatedly managed this, not even in five attempts.

A similar pattern repeats elsewhere. For a real bug in a popular package manager, the model found the cause and fixed an edge case that a community patch overlooked. According to Anthropic, a competing model only treated the symptom and declared the bug resolved. The new feature is therefore called self-verification – instead of submitting the first result, the model goes back and checks if something actually works.

Fewer restrictions than the more expensive model

A practical reason why Opus 5 might be more appealing than Fable 5 lies beyond performance. Fable and Mythos are subject to a thirty-day data retention policy, which bothers some users – Opus 5, like its predecessor, does not fall under this regime. Additionally, it has significantly looser security classifiers: Anthropic estimates that for Opus 5, they will intervene approximately 85% less often than for Fable.

But the limitations have not disappeared. The model is allowed to search for vulnerabilities in source code; scanning binaries, penetration testing, and exploit creation remain blocked. However, a blocked query no longer has to return as an error – in the beta feature Automatic Fallbacks, it is redirected to a weaker model that will respond to it.

The release pace is noteworthy in itself. Opus 4.8 was released on May 28th, Mythos 5, Fable 5, and Sonnet 5 in June, and Opus 5 arrives in July. In less than two months, Anthropic has thus refreshed virtually its entire portfolio, and the only model still awaiting the Series 5 generation remains the lightweight Haiku.

Would you pay for the most expensive model when the cheaper one can do the same?

Source: Anthropic