Anthropic introduces Claude Fable 5.1 for coding and AI

Anthropic launched Claude Fable 5.1, an AI model designed for programming, research and knowledge work, alongside Claude Mythos 5.1, a less restricted version for authorized professionals in cybersecurity and the life sciences.

The two models are technically identical. The difference is in their safeguards: Fable 5.1 is generally available, while Mythos 5.1 is offered only through verified-access programs.

More capability at a lower cost

Anthropic says Fable 5.1 outperforms its predecessor in programming, computer use, reasoning and tasks that require extended work. In Terminal-Bench 4.0, an agentic programming benchmark, it scored 55.8%, compared with 42% for Fable 5.

It also improved on other tests:

• 52.6% in agentic scientific research, compared with 24.7% for Fable 5.
• 77.9% on the partial version of OSWorld 2.0, which measures computer use.
• 31.4% on AutomationBench, focused on enterprise workflows, compared with the previous 17.1%.
• 65% on Humanity's Last Exam with tools.

These figures come from Anthropic's published evaluations and do not mean the model is always correct. They do point to a particularly relevant improvement for tasks in which AI must research, execute steps and correct its own mistakes.

The company is also lowering the price of cached context reads, meaning situations in which the model reuses information it has already processed. That cost is falling by 75%, to $0.25 per million tokens. For typical workloads, Anthropic estimates a reduction of around 25% compared with Fable 5. For complex, highly autonomous programming tasks, the savings can reach approximately 45%.

The rest of the pricing remains at $10 per million input tokens and $50 per million output tokens.

What it can do in science

Anthropic tested Mythos 5.1 with protein-design tools and sent its results to external organizations for experimental validation. According to the company, the model achieved a viable-design rate close to 50% across 12 targets, compared with the typical 10% to 15% for this type of design.

For three specific targets, its best designs had binding affinities up to ten times higher than the best results submitted in Adaptyv Bio competitions. These are results from specific tests, not evidence that the model can create drugs on its own.

Fable 5.1 also trained a neural network to produce an elevation map of one-third of Venus from radar images captured by NASA's Magellan mission. The new map provides details at a scale of between two and three kilometers, compared with the previous 10 or 20 kilometers, according to Anthropic.

Mythos 5.1, meanwhile, optimized seven deep-learning models used in biology with GPU-specific code and by reusing intermediate results. The company says it accelerated them by up to 2.5 times, which could reduce GPU costs by between 30% and 60% for certain analyses.

Fewer cybersecurity blocks, with limits

One of Fable 5.1's practical new capabilities is identifying software vulnerabilities to help fix them. Anthropic says its safety systems generate 60% fewer false positives in cybersecurity, meaning they block fewer legitimate requests than before.

That does not cover every security task. The model still refers dual-use activities to other systems, including penetration testing, exploit generation and certain binary-based vulnerability analyses.

In biology, medical and basic-biology questions trigger the filters 85% less often than when Fable 5 launched. Life sciences research and development queries remain restricted and are directed to Opus models, while advanced access to Mythos 5.1 is reserved for verified professionals.

Mythos 5.1 is not for everyone

Claude Mythos 5.1 is distributed through two programs:

• The Cyber Verification Program, for defensive cybersecurity work.
• The Life Sciences Verification Program, created for research and development professionals in the life sciences.

For now, access is limited to a group of US organizations. Anthropic says it is working with the US government to expand access to more domestic and international partners.

The company also acknowledges limits in its safety testing. Its evaluations provide less coverage for tasks with very long contexts, multiple agents and impossible requests. The model can also occasionally bypass approvals and automated classifiers.

Enterprise privacy and watermarking

Anthropic is preparing a system called Enterprise Frontier Safeguards. It will allow a company's data to be stored in its own cloud infrastructure rather than in Anthropic's systems, with privacy equivalent to a zero-retention policy.

Deployment will begin in phases starting in fall 2026. Until then, eligible customers will be able to use Fable 5.1 with zero data retention.

Models released after August 2, 2026, will also include an invisible watermark in their text, as required by the transparency framework of the European Union's AI Act. It contains no information about the user and does not change the generated content. Anthropic is testing an API that will allow authorized organizations to check whether a text probably contains the watermark.

Claude Fable 5.1 is already available in Claude and on the Amazon Web Services, Google Cloud and Microsoft Azure platforms. For you, the most important change will depend on how you use it: programming and task automation may become cheaper, while advanced biological research and some cybersecurity operations will remain under controlled access. The question to watch is whether these improvements maintain their performance when they move from Anthropic's tests to real projects, with data, errors and risks that are difficult to anticipate.