Anthropic has released Claude Fable 5.1 for general use and Claude Mythos 5.1 for vetted organizations. Both use the same underlying model, but Mythos comes with fewer restrictions for advanced cybersecurity and life-science research.
Anthropic is starting September with one of its biggest Claude releases of the year.
The company has launched Claude Fable 5.1, its latest model for coding, reasoning and professional knowledge work, alongside Claude Mythos 5.1, a more restricted-access version of the same underlying model designed for organizations working on advanced cybersecurity and biological research.
The distinction matters.
Anthropic is effectively creating two paths for the same model. Fable 5.1 is the broadly available version, with production safeguards enabled. Mythos 5.1 is reserved for vetted organizations that need access to capabilities that are considered more sensitive.
That makes this more than another incremental model update.
Anthropic is openly acknowledging that some AI capabilities have become powerful enough that who gets access can matter almost as much as the model itself.
Fable 5.1 Is Built for Long-Running Work
Anthropic is targeting Fable 5.1 at tasks that require more than a single prompt and response.
The model supports a 1-million-token context window and can generate up to 128,000 tokens in a single response. Base pricing remains $10 per million input tokens and $50 per million output tokens.
The more interesting pricing change is cache.
Cache reads are now priced at $0.25 per million tokens, a 75% reduction. Anthropic estimates that the change can reduce real-world costs by about 25% for typical workloads and as much as 45% for highly agentic workloads.
For developers running agents that repeatedly work with large amounts of context, that could have a meaningful impact on operating costs.
The model can retain and reuse previously processed context instead of effectively paying the same price for it over and over.
That matters as AI systems move from chat interfaces toward agents that spend much longer working on a task.
The Benchmark Gains Are Significant
Anthropic reports major improvements across several of its evaluations.
On Terminal-Bench Science 0.1, Fable 5.1 scored 52.6%, compared with 24.7% for Fable 5. On Terminal-Bench 4.0, it reached 55.8%, compared with 42.0% for the previous model.
These benchmarks focus on more complex workloads involving coding, tool use, reasoning and multi-step tasks.
Anthropic says Fable 5.1 is particularly suited to agentic coding, scientific research and long-running professional workflows, including documents, spreadsheets and presentations.
That is the direction the company is clearly pushing.
Claude is being positioned less as a chatbot that answers questions and more as software that can take a goal and work through it over many steps.
Mythos 5.1 Is the Same Model With Different Safeguards
The most unusual part of the launch is Claude Mythos 5.1.
Anthropic says Fable 5.1 and Mythos 5.1 share the same underlying model. What changes is the level of safeguards and the organizations allowed to use the more permissive configuration.
Fable is the general-purpose version.
Mythos is intended for vetted users working in areas such as cybersecurity and life sciences.
Anthropic’s reasoning is that a model capable of advanced work in these areas can also create serious risks if its capabilities are misused.
That is why Mythos 5.1 is not being opened to everyone.
Access is tied to Anthropic’s trusted programs, with separate pathways for cybersecurity and life-science research.
The model itself is therefore becoming part of the access-control equation.
Instead of asking only whether an AI system is safe enough to release, companies increasingly have to decide which capabilities should be available to which users.
Anthropic Is Also Trying to Make Claude Less Frustrating
Fable 5.1 is not simply more capable.
Anthropic has changed its safety systems as well.
The company says its cybersecurity safeguards now produce around 60% fewer false positives in Claude Code sessions. That means legitimate requests involving security research should be less likely to trigger unnecessary refusals.
Anthropic says Fable 5.1 can help users discover software vulnerabilities, while still blocking requests to develop exploits for those vulnerabilities.
It is a deliberately narrow distinction.
Security researchers need AI systems that can inspect software, understand vulnerabilities and help identify weaknesses.
At the same time, giving an AI model detailed instructions for turning those weaknesses into working attacks introduces a very different risk.
Anthropic is trying to keep Fable on the defensive side of that line.
The company is also introducing invisible watermarking for Claude-generated text and files, alongside a detection API being tested with selected organizations.
That could become increasingly relevant as AI-generated material becomes harder to distinguish from human work.
One Model, Two Doors
The Fable 5.1 and Mythos 5.1 launch points toward a model of AI distribution that could become much more common.
Anthropic is not making its most capable configuration universally available.
Instead, it is offering a general version with stronger safeguards while allowing vetted specialists to access a more permissive version when there is a legitimate reason to need it.
For most users, Fable 5.1 is the important release: stronger coding and research, a 1-million-token context window and substantially cheaper cache reads for long-running workloads.
For cybersecurity teams and researchers, Mythos 5.1 is the more significant development.
It shows that AI companies are increasingly treating capability itself as something that needs to be controlled.
That could become one of the defining issues of the next generation of AI models.
The question is no longer simply how intelligent the model is.
It is also who gets to use its most powerful abilities, under what conditions, and with how much oversight.
Anthropic has made its decision for now.
The most capable version of Claude is not open to everyone.
by The Tech Spot Editorial Team
