Three months after the first Mythos model for the general public, Anthropic is following up. Fable 5.1 replaces the previous top-of-the-line model, is about a quarter cheaper for typical applications – and is the first model from the company to bear a watermark based on European law.
In June, Anthropic released Fable 5, the first Mythos-class game to be made generally available. Its successors, Fable 5.1 and Mythos 5.1, are now available, and as with the first generation, they are the same model with varying levels of protection. Details can be found in Anthropic's announcement and the accompanying system card.
Key Facts at a Glance
- Fable 5.1 is now available on all platforms via the programming interface under the name claude-fable-5-1.
- The prices for inputs and outputs remain unchanged; the model becomes cheaper through the costs for cached content.
- For typical workloads, total costs decrease by around 25 percent, and by up to 45 percent for work heavily supported by tools.
- The security filters are triggered less frequently for harmless requests – around 60 percent less often in the security area.
- Fable 5.1 is the first model released after August 2nd to feature an invisible text watermark, as required by the EU's code of conduct on artificial intelligence.
Where the model gains ground
Anthropic identifies three main areas: programming, knowledge-intensive work, and tasks that extend over long periods. The most striking improvement is seen in a test procedure for scientific work in the terminal, where the result more than doubled compared to the previous version.
| Test procedure | Fable 5.1 | Fable 5 | Opus 5 |
|---|---|---|---|
| Scientific work (Terminal-Bench-Science 0.1) | 52,6 % | 24,7 % | 29,0 % |
| Programming in the terminal (Terminal-Bench 4.0) | 55,8 % | 42,0 % | 52,3 % |
| Computer-aided operation (OSWorld 2.0, strict) | 41,7 % | 36,1 % | 39,6 % |
| Interdisciplinary reasoning (Humanity's Last Exam, without tools) | 60,9 % | 57,8 % | 56,6 % |
| Business processes (AutomationBench) | 31,4 % | 17,1 % | 26,9 % |
A remarkable example from practical testing: At a financial institution, the model identified the cause of a crash that occurred approximately once per million runs and which no one there had been able to explain for years. To do this, it disassembled an external program library and compared it to the memory image.
It will be cheaper at a certain point
The prices for input and output tokens remain unchanged at $10 and $50 per million, respectively. Anthropic has reduced the cost of rereading previously processed content by 75 percent to 25 cents per million tokens.
This may sound like a minor detail, but in practice it makes all the difference. In workflows where a model repeatedly accesses the same context, these accesses account for the majority of the costs. Anthropic estimates the savings at around a quarter for typical applications and up to 45 percent for highly tool-driven work.
Fewer false alarms in the protection mechanisms
According to the company, the filters that intercept sensitive requests are now more precise. In the area of cybersecurity, they are said to trigger around 60 percent less often for harmless requests, and 85 percent less often for simple biological and medical questions than at the launch of Fable 5.
In terms of functionality, one boundary has changed: Fable 5.1 is now permitted to detect security vulnerabilities in software, but not to develop attack tools for them. Tasks with dual purposes – penetration tests, the creation of exploits, and scanning program files for vulnerabilities – remain the responsibility of the smaller Opus models.
Also new is a block against the appropriation of skills by external providers: Newly created developer accounts can no longer retrospectively edit the model's previous thought process in a conversation.
The watermark comes from Brussels
For German-speaking countries, another paragraph of the announcement is the most interesting. In July, Anthropic, along with 190 other signatories, signed the Code of Conduct on the Transparency of AI-Generated Content, which is part of the EU law on artificial intelligence. This results in a concrete obligation: models published after August 2, 2026, must watermark their text output.
Fable 5.1 is the company's first model to implement this feature. The watermark is a numerical method that indicates the probability of a text originating from Claude. It is invisible without the appropriate interface, does not alter the output, and, according to the company, contains no information about users or their conversations.
The associated detection interface is running in a closed preview. Initially, access will be granted to the groups stipulated by EU law – authorities, law enforcement, media, fact-checkers, independent research institutions, educational institutions, and civil society organizations within the EU. For newsrooms in Germany, this is the most practically relevant development of the day: For the first time, there is an official way to verify suspected machine-generated text, instead of relying on detection services whose accuracy is regularly criticized.
Mythos 5.1 will remain in the USA for now
The more permissive version is still not available on the local market. Mythos 5.1 is exclusively available to vetted organizations, and initially only in the United States. Anthropic states that they are working with the US government to expand the network to include other domestic and international partners.
This continues a pattern that has accompanied the Mythos series from the beginning: first the suspension in June by order of the US Department of Commerce, then the lifting of the export ban at the end of June, and the return on July 1st. Those who work in Germany will continue to receive the protected version.
Price becomes the main argument
What's striking about this announcement is where Anthropic is placing its emphasis. A model that performs well in testing is available weekly; more interesting is the shift in pricing. Those who previously opted for smaller models due to price can now get top-of-the-line performance at a price closer to the previous mid-range offering – a partner company is announcing precisely this change for code verification.
For users in German-speaking countries, everyday life will initially remain unaffected, because the price change is tied to token-based billing and not to subscription models. It will become noticeable where companies build their own tools on this model – and there it is likely to cause some significant changes.
- Apple: Leadership reshuffle continues on the first day
- John Ternus: First day as Apple CEO with a clear message
Do you use AI models professionally, and would a quarter less cost influence your choice – or does the only thing that matters to you in the end is which model solves the task best?