Anthropic released Claude Opus 5.5. In tests, it surpasses Fable 5.1 and is expected to be much cheaper

Less than two months after the release of Opus 5, Anthropic has its successor. Claude Opus 5.5 opens a new 5.5 series, and the company positions it as a cheaper path to the performance of the flagship Fable 5.1. If you work with Claude via API, regular work is expected to cost you about 40% less than with Opus 5. Subscribers, in turn, will receive more generous limits.

Opus 5.5 surpasses Fable 5.1 in tests

In its announcement, Anthropic states that the new model leads in autonomous programming, computer control, and knowledge work. In the Terminal-Bench 4.0 test, which measures the handling of complex command-line tasks, it reports 66.4% compared to 55.8% for Fable 5.1 and 52.3% for Opus 5. According to numbers published by OpenAI, the competing GPT-6 Astra achieved 57.9%.

However, all results come from tests conducted by the company itself, and we cannot independently verify them yet. Anthropic, in fact, admits that differences in benchmark tests are less reliable indicators of real-world practice. According to them, the lead over Fable 5.1 in daily work is smaller than what the tables suggest.

Specific examples are more illustrative. One tester reportedly reviewed and fixed 200,000 lines of code in less than three hours with the new model, while Opus 5 needed over 20 hours for the same task. In an internal test, Opus 5.5 then rewrote HAProxy, a widely used software for distributing web traffic among servers, from C to Rust in 9.5 hours. Fable 5.1 took 12 hours and cost roughly twice as much.

The way the model writes has also changed. Anthropic is responding to one of the most frequent criticisms of Opus 5, which concerned the clarity of its responses. The new model is designed to provide the most important information upfront, use less jargon, and better adhere to the writing rules you provide. Additionally, it generates text more than 30% faster.

The Price is More Favorable

Developers will pay approximately 85 CZK (4 dollars) per million input tokens and 425 CZK (20 dollars) per million output tokens, which is one-fifth less than with Opus 5. Cache reading, which the company states makes up most of the costs for programming and agent work, has become significantly cheaper – from less than 11 CZK to approximately 4 CZK per million tokens. Because the model also consumes fewer tokens for the same task, Anthropic estimates the savings in typical use to be the aforementioned 40%.

Subscribers to Pro, Max, Team, and corporate Enterprise plans with per-user licenses will receive higher five-hour usage limits. In addition, Anthropic is adding a one-time limit reset that you can save and apply when you need it – for example, when your quota runs out in the middle of ongoing work.

Opus 5.5 is available now on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Sonnet 5.5 and Haiku 5.5 are expected to follow in the coming weeks.

Why Did It Receive Stricter Safeguards?

This is the first model since Anthropic CEO Dario Amodei called last week for AI development to proceed only as quickly as safety measures allow. Before its release, it was also tested by external evaluators, including METR and Frontier Design. In its own behavior audit, which tests the model in nearly 2,000 simulated situations, the company states that it achieved the best result of all models tested to date.

However, in biology and cybersecurity, according to Anthropic, it matches the limited-availability Mythos 5.1, and thus received similar safeguards to Fable 5.1. While it will still be possible to find and fix errors in its own code, most cybersecurity tasks will be automatically redirected to the older Opus 4.8. This is a noticeable change compared to Opus 5, whose filters were significantly looser. Verified laboratories and pharmaceutical companies can apply for looser access in a new biological research program; a similar program for security experts will be extended to Opus 5.5 in the coming weeks.

The company also admits that during tests, the model often senses it is being tested, which makes it difficult to estimate how it will behave in real-world deployment. Opus 5.5 also can no longer be used with “thinking” turned off, and like Fable 5.1, it marks texts with an invisible watermark due to the European AI Act, about which we wrote in August. Conversely, like previous Opuses, it is also available in a data-retention-free mode.

Will you opt for Opus 5.5, or stick with the more expensive Fable 5.1?

Sources: Anthropic, X