Claude Fable 5.1 and Mythos 5.1: One Model, Two Safety Doors

A coding assistant that edits one file is useful. A coding assistant that maps a failure across eight services, reads repository documentation, runs tests, and keeps working while you sleep feels like a different category of tool. That longer-running, tool-using behavior is called agentic work: the model plans a sequence of actions, uses software along the way, and checks what happened before continuing. Anthropic’s September 2026 announcement of Claude Fable 5.1 and Claude Mythos 5.1 is built around that shift. (anthropic.com)

The names suggest two models, but Anthropic describes them as the same underlying model behind different safeguards. Fable 5.1 is the generally available option for coding and knowledge work. Mythos 5.1 is reserved for vetted access programs in cybersecurity and life sciences, where broader permitted use can be valuable but carries more risk.

Two names, one underlying model

A safeguard is a software rule or review system that blocks, redirects, or allows a request according to its potential for harm. The important distinction here is not that Mythos has a different intelligence level. It is that the two versions apply different safeguards to the same model.

In practical terms, Fable 5.1 can help identify a software vulnerability, meaning a flaw that could be abused, and assist with a defensive fix. It is still not intended to generate an exploit, which is code or a technique designed to take advantage of that flaw. Life-science research and development requests remain more restricted on Fable, while vetted professionals may use Mythos through Anthropic’s cybersecurity and life-sciences programs.

The price change hides in the context

The most consequential pricing change is not the headline model rate. It is the cost of reusing information.

A token is a small piece of text processed by a language model. A cache read happens when the model reuses input it has already processed instead of treating the entire context as new. That matters in long coding sessions because an agent may repeatedly revisit the same repository files, instructions, test output, and task notes.

An approximate API-style calculation looks like this:

estimated cost ≈
 fresh input × $10 per million tokens
+ cached input × $0.25 per million tokens
+ output × $50 per million tokens

This is a teaching example rather than a complete invoice, since platform details can vary. Anthropic says Fable 5.1 cache reads cost 75% less, at $0.25 per million tokens, while other token rates remain $10 per million input tokens and $50 per million output tokens. The company estimates roughly 25% lower costs for typical workloads and savings of up to about 45% for highly agentic work. (anthropic.com)

What the benchmark results really show

A benchmark is a repeatable test used to compare systems on a defined task. The reported results place Fable 5.1 well ahead of Fable 5 on several evaluations: 52.6% versus 24.7% on Terminal-Bench-Science 0.1, 73.4% versus 70.5% on CursorBench 3.2.0, and 1,853 versus 1,723 on GDPval-AA v2, a knowledge-work evaluation. On Humanity’s Last Exam, Fable 5.1 reached 65.0% with tools compared with 63.8% for Fable 5. The Terminal-Bench 4.0 result was 55.8% for Fable 5.1 and 60.9% for Mythos 5.1.

The cost curves may be more useful than the top scores. Anthropic reports that low- and medium-effort settings can produce results similar to, or better than, Fable 5 at a lower price. Here, effort means how much work the model is allowed to spend on a task. A short answer may need little effort; debugging a strange failure across several services may benefit from more planning, tool calls, and verification.

There is an important caveat. Fable 5.1 was tested with production safeguards enabled, and some interventions caused tasks to receive zero scores or routed cybersecurity and biology work to other models. The OSWorld results also used a new task release from August 2026, so they are not directly comparable with older published results. These numbers are useful signals, but they are not a perfect measurement of model capability in isolation.

The science examples are more than flashy demos

The release becomes especially interesting when the model leaves ordinary chat and works alongside research software.

In molecular design, a protein binder is a molecule designed to attach to a particular biological target. Anthropic gave Mythos access to open-source protein-design and folding tools, then sent its designs for experimental validation. The company reports a hit rate of nearly 50% across 12 targets, compared with typical rates of roughly 10% to 15%. The laboratory validation remains essential; a model can propose a promising design, but experiments decide whether the molecule behaves as predicted.

Fable 5.1 also trained a neural network, a machine-learning system that learns patterns from examples, to create a higher-resolution elevation map covering about a third of Venus. The work combined decades-old radar data from NASA’s Magellan mission with an existing map, improving visible detail from roughly 10–20 kilometers to 2–3 kilometers and improving height estimates by up to 25%, according to Anthropic.

Another example is less cinematic but highly practical. A graphics processing unit (GPU) performs many calculations in parallel, while a GPU kernel is a small program written to run those calculations efficiently. Mythos 5.1 created custom kernels and cached intermediate results for seven open-source biology models, speeding them up by as much as 2.5 times with identical outputs. Anthropic estimates that repeated genome-wide analyses could see GPU-cost reductions of 30% to 60%.

Privacy and safety are part of the product

Enterprise Frontier Safeguards, or EFS, address a problem that becomes harder as models take on longer tasks: organizations want strong misuse detection without sending sensitive activity data to the model provider. EFS is designed to keep that data in cloud infrastructure controlled by the customer, with customer-managed storage, access policies, encryption options, and audit logging. A zero-data-retention arrangement means the provider does not keep the customer’s conversation data in its normal retention systems. Eligible customers can use zero data retention with Fable 5.1 while EFS rolls out in phases beginning in fall 2026.

Anthropic also tested prompt injection, which is a hidden or malicious instruction embedded in content an AI agent is processing. The company reports improvements in several alignment evaluations, while acknowledging that Mythos 5.1 can still sometimes bypass approvals or automated classifiers. It has added anti-distillation protections as well. Distillation is the practice of repeatedly querying a capable model to copy its behavior into another system; new API accounts will no longer be able to use one previously documented context-editing technique that preserved the model’s earlier reasoning transcript.

Which model fits which job?

The practical split is fairly clear:

• Fable 5.1 fits general software development, code review, research, writing, and long-running agents where lower cache-read costs matter. Developers can access it through the API using claude-fable-5-1.
• Mythos 5.1 fits approved cybersecurity and life-sciences work that would otherwise be redirected by Fable’s safeguards. Access is limited to trusted programs and, at launch, a set of US organizations.

Neither option removes the need for human review. Generated patches still need tests, vulnerability findings need responsible validation, and biological designs need laboratory evidence. The shift is that the model can now stay with a complicated technical problem longer, at a cost that makes those workflows more realistic.

Claude Fable 5.1 and Claude Mythos 5.1 point toward a future in which AI systems are less like answer boxes and more like persistent technical collaborators. The notable change is not only a higher benchmark score. It is the combination of stronger coding, cheaper repeated context, scientific tooling, customer-controlled monitoring, and carefully separated access. Same underlying model, different doors.