Anthropic has introduced Claude Fable 5.1 and Claude Mythos 5.1, two versions of the same model aimed at coding, agents, professional work, and scientific research, but differentiated by their safety controls. Fable 5.1 is generally available, while Mythos 5.1 is initially restricted to verified cybersecurity and life-sciences professionals.
Key facts about Claude Fable 5.1 in 30 seconds
- Fable 5.1 improves especially at agentic coding, scientific research, automation, and computer use.
- Anthropic estimates costs about 25% lower for typical workloads and up to 45% lower for context- and tool-intensive processes.
- Cache reads drop 75%, to $0.25 per million tokens.
- Mythos 5.1 shares the same model as Fable, but uses different controls for verified professionals.
- Fable 5.1 can search for software vulnerabilities, though it keeps restrictions on generating exploits and other dual-use activities.
The launch also shows where competition among the most advanced models is shifting. Comparisons no longer depend solely on answering a question correctly or generating code on a test. Anthropic is focusing on long-running tasks, tool use, research, computer interaction, and, especially, how much it costs to keep an agent working over extended periods.
That last point explains one of the seemingly small changes in Fable 5.1 that could end up carrying real weight in production: the cut in cache pricing.
Fable 5.1 Improves Most When AI Has to Work for Longer
Anthropic presents Fable 5.1 as a significant evolution from Fable 5 and backs that up with results across several tests.
On Terminal-Bench-Science 0.1, designed to evaluate agentic scientific research, Fable 5.1 scores 52.6%, compared with 24.7% for Fable 5. In Anthropic’s own tests, Opus 5 scores 29.0% and GPT-5.6 Sol 22.4%.
In agentic coding via Terminal-Bench 4.0, Fable 5.1 scores 55.8%. Mythos 5.1 reaches 60.9%, while Fable 5 scores 42.0% and Opus 5 scores 52.3%.
The gaps are smaller on other tests, but they still show improvements over the previous generation.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | 55.8%* | 42.0% | 52.3% | 37.3% |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
* Mythos 5.1 reaches 60.9% on Terminal-Bench 4.0.
The results come from evaluations published by Anthropic and should be read within the configurations and methodology the company describes. Not every model was tested on every benchmark, either.
More interesting than a gap of a few points may be how the model behaves during long-running processes.
Today’s coding agents can open files, modify code, run tests, consult documentation, and revisit their own changes. The longer the process runs, the greater the chance the model loses track of the problem, takes a wrong turn, or ends up applying a superficial fix.
Anthropic says Fable 5.1 shows a greater ability to trace a problem back to its root cause. As an example, it cites Millennium, where the model reportedly identified the source of a rare bug in its internal systems that had gone unexplained for several years.
It’s a case supplied by an early-access customer, not an independent benchmark, but it illustrates well the kind of problem these models are being designed to handle.
Cache Is Becoming an Economic Piece of the Agent Puzzle
Fable 5.1 keeps Fable 5’s headline pricing: $10 per million input tokens and $50 per million output tokens.
The difference shows up when the model reuses previously processed context.
Anthropic is cutting the price of cache reads by 75%, down to $0.25 per million tokens.
That might look like a minor tweak next to input and output pricing, but agents reuse huge amounts of information.
A coding agent can keep repository files, documentation, instructions, previous outputs, conversations, and tool responses in context. If it needs to repeatedly consult that information over dozens of steps, avoiding the full price for reprocessing it changes the total cost of the task.
Based on four weeks of usage Anthropic observed in August 2026, Fable 5.1 works out roughly 25% cheaper than Fable 5 for a typical workload.
For highly agentic workloads that are heavy on context and tool use, the company estimates the reduction can approach 45%.
That doesn’t mean using Fable 5.1 will automatically cost 45% less. The savings will depend on how much context can be reused and on the split between input, output, and cache tokens.
But it points to a question that’s becoming increasingly important: evaluating the economics of a model used as an agent can no longer rely on just the nominal price per million tokens.
Fable Can Find Vulnerabilities; Mythos Widens Defensive Access
Cybersecurity is another major difference in this generation.
Anthropic will allow Claude Fable 5.1 to be used to identify software vulnerabilities, widening the defensive tasks that can be carried out without triggering system controls.
The company says it has made its safeguards more precise and estimates Claude Code users will see around 60% fewer security interventions per session compared with Fable 5’s previous protections.
This doesn’t remove the restrictions.
Certain dual-use activities, including exploit generation, some penetration testing, and binary vulnerability analysis, remain subject to additional controls.
For work that needs more freedom, there’s Claude Mythos 5.1.
Anthropic says Fable 5.1 and Mythos 5.1 are the same model, but Mythos uses safeguards tailored to previously verified professionals.
Access is managed through the Cyber Verification Program and the Life Sciences Verification Program. Mythos 5.1 is initially available to a limited set of U.S. organizations, though Anthropic plans to expand the program later.
Claude Security, the company’s tool that scans repositories to find vulnerabilities and proposes patches later subject to human review, now runs on Mythos 5.1.
This split between technical capability and usage permissions could become a common feature of frontier models. The same model can behave differently depending on the user’s identity and purpose, especially in dual-use areas.
Anthropic Wants to Bring Agents to Scientific Research Too
The other major area of the announcement is science.
Anthropic tested Fable 5.1 and Mythos 5.1 on molecular design, scientific analysis, and optimizing models used in computational biology.
In one experiment, Mythos 5.1 generated binder molecule designs that were later subjected to external experimental validation. Anthropic says that on three targets it achieved affinities ten times higher than the best designs from Adaptyv Bio competitions, and that the rate of viable designs came close to 50% across 12 targets.
In another project, Fable 5.1 trained a neural network to build an elevation map of roughly a third of Venus using radar imagery captured by NASA’s Magellan mission more than three decades ago.
The resulting map’s resolution reaches, according to Anthropic, detail levels of two to three kilometers, compared with 10-20 kilometers in earlier work, and improves height accuracy by up to 25%.
Mythos 5.1 also wrote custom GPU kernels and applied caching techniques to speed up seven open deep-learning models used in genomics and protein research. The published gains range from 1.4x to 2.5x over an NVIDIA H100.
Anthropic estimates these optimizations could cut the GPU cost of certain analyses by 30% to 60%.
These are specific experiments selected by the company itself, but they show that the goal for these models goes beyond a conventional chatbot: take in a problem, use specialized software, produce results, and sustain considerably longer workflows.
Enterprise Privacy and an Invisible Watermark for Text
Anthropic is also introducing Enterprise Frontier Safeguards (EFS), a system designed to combine misuse controls with enterprise privacy requirements.
With EFS, data will be stored on cloud infrastructure controlled by the customer rather than on Anthropic’s systems. The company equates the resulting privacy level with a zero-retention agreement.
Rollout will begin in phases this fall. Anthropic says it has worked on the system with more than 100 customers and plans to support it across Claude Code, Claude Enterprise, Claude Platform, and various cloud platforms.
Fable 5.1 also includes another unusual feature: an invisible watermark in generated text.
Anthropic ties the move to its commitments under the Code of Practice on Transparency of AI-Generated Content tied to the EU’s AI Act.
The signal doesn’t appear visually in the document. It’s designed to let an API estimate whether Claude took part in generating the content, without including information about the user, their organization, or their conversations.
The detection API is starting in private preview for certain organizations, including regulators, media outlets, fact-checkers, researchers, and entities that need to check for this kind of signal.
Fable 5.1 is available through Anthropic’s platforms and as claude-fable-5-1 in its API. Mythos 5.1 will initially remain behind controlled-access programs.
Beyond the specific benchmark results, the launch highlights three variables gaining weight in the next generation of models: how long they can work autonomously, how much it costs to sustain that work, and what capabilities can be offered without granting every user the same level of access.
Frequently asked questions
What’s the difference between Claude Fable 5.1 and Mythos 5.1?
Anthropic says they use the same model. The main difference is in the safeguards: Fable 5.1 has general availability, while Mythos 5.1 offers more permissive controls for verified cybersecurity and life-sciences professionals.
How much does Claude Fable 5.1 cost?
The base price is $10 per million input tokens and $50 per million output tokens. Cache reads cost $0.25 per million, 75% less than before.
Can Fable 5.1 be used to search for vulnerabilities?
Yes. Anthropic allows it to be used to identify software vulnerabilities for defensive purposes. Activities such as exploit generation and certain penetration tests remain subject to additional restrictions.
Why can Fable 5.1 be cheaper for agentic tasks?
Agents repeatedly reuse large amounts of context. By cutting the price of cache reads, Anthropic estimates savings of close to 25% on typical workloads and up to roughly 45% on processes that are very heavy on context and tool use.
