Anthropic has released Claude Fable 5.1 and Mythos 5.1, introducing lower pricing and major performance improvements for developers and researchers. The company says cache read pricing has fallen 75%, while its latest model more than doubles the score of its predecessor on a scientific research benchmark.
Fable 5.1 is the generally available model, while Mythos 5.1 is being offered to vetted users through specialized access programmes. Those programmes currently focus on cybersecurity and life sciences, with Anthropic saying its biology programme was developed alongside the US government.
The biggest developer change is pricing. Cache reads now cost $0.25 per million tokens, down 75% from the previous price. Anthropic says the reduction could lower typical workload costs by around 25%, while highly agentic workloads could see savings of up to 45%.
Performance has also taken a significant step forward. Anthropic reports that Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1, compared with 24.7% for Fable 5. The new model also recorded a 55.8% score on Terminal-Bench 4.0.
The pricing changes matter beyond ordinary chatbot usage. AI agents can generate significantly more model activity because they may repeatedly read context, call tools and execute multiple steps before completing a task.
Lower cache read costs therefore make longer and more complex agentic workflows easier to run economically. If agents become more common in software development, research and enterprise operations, reducing the cost of repeatedly accessing stored context could become a major competitive advantage.
Anthropic is also positioning Mythos 5.1 for highly specialized work. Access remains restricted, with the company currently limiting the model to vetted organizations involved in cybersecurity and life sciences.
That approach reflects Anthropic’s broader focus on deploying increasingly capable models while maintaining tighter controls around their most sensitive capabilities.
Claude now carries an EU AI Act watermark
The release also comes with an important European development. Anthropic says outputs from models released after August 2, 2026 carry an invisible watermark designed to satisfy transparency requirements under the EU AI Act.
The watermark does not reveal information about the user and cannot be seen without a detection tool. Anthropic signed the EU’s Code of Practice on transparency for AI-generated content alongside other major organizations earlier this year.
The detection system, however, is not as broadly available as previously expected. Anthropic is currently offering its detection API through a private preview for eligible organizations, including regulators, law enforcement, media organizations, fact-checkers, researchers and EU civil society groups.
That creates an interesting gap between the regulation and the technology used to enforce it. The EU requires synthetic content to carry machine-readable markings that can be detected, but access to Anthropic’s verification tool remains restricted.
Anthropic says enterprises with their own compliance obligations can also apply for access and that it plans to expand availability over time.
The launch shows where competition between AI companies is heading. Model quality still matters, but the economics of running increasingly autonomous systems may become just as important.
A model that performs well on benchmarks but costs too much to operate can struggle when deployed across thousands of agentic workflows. Cutting cache read prices by 75% gives Anthropic more room to compete in that market while its newer models continue improving on technical benchmarks.
At the same time, Anthropic is entering a new phase of AI regulation in Europe. Its models are becoming cheaper and more capable, while regulators are demanding greater transparency around how AI-generated content can be identified.
The technology is moving quickly. The infrastructure and rules around it are still catching up.
