Anthropic announces Claude Fable 5.1 and Mythos 5.1

Fable and Mythos share a model but differ in safeguards

The different names may suggest that Mythos 5.1 is a larger model. But Anthropic says they are the same underneath. Fable has restrictions for broad availability, while Mythos adjusts those restrictions for vetted users conducting cyber defense or life-science research.

This distinction matters when choosing a model. You do not need Mythos access simply to write text or fix code. In research and security, however, a model's ability to solve a problem and its willingness to accept a request are separate questions.

Mythos access involves checks on the organization and intended use. Initially, it is limited to selected US organizations, with international expansion planned later. It should not be assumed to be available from Japan on the same terms as Fable under an ordinary subscription.

Gains over the predecessor in coding and scientific research

In the developer's comparison table, Fable 5.1 scores 55.8% on Terminal-Bench 4.0, which tests coding work through a terminal, versus 42.0% for Fable 5 and 52.3% for Opus 5. On Terminal-Bench-Science 0.1, covering scientific research tasks, its score rises from 24.7% to 52.6%.

These are Anthropic's measurements in its own environment. Some tasks that triggered safeguards were handed over to another Opus model. The scores therefore reflect restrictions and the execution environment as well as the model's own capabilities.

A 55.8% score on a coding benchmark is not the success rate for your work. For code changes, comparisons should also consider whether changes stay focused, tests pass and the amount of subsequent human repair decreases.

The price cut is for rereading previously processed input

API pricing remains $10 per million input tokens and $50 per million output tokens, unchanged from Fable 5. Cache reads drop from $1 to $0.25, a 75% reduction in the rate for reusing stored input.

Tasks that read long codebases or documents and repeatedly continue work with tools often refer back to the same input. If those parts can use the cache, the price cut can reduce the whole task's cost. A short, one-off question that produces a long answer benefits much less.

For example, reading one million cached tokens and producing 10,000 output tokens costs $0.75 instead of $1.50. This is the editors' calculation from the price table, excluding the initial cache write. With 100,000 output tokens, the cost falls from $6 to $5.25, a reduction of only 12.5%.

A cache-read rate one quarter as high does not mean a total bill one quarter as high. Separate cache-writing costs, successfully reused input and generated output to see what the price cut means for your workload.

The cost comparison with Opus depends on input reuse

In the same pricing table on the announcement date, Opus 5 costs $5 for input, $25 for output and $0.50 for cache reads. Fable 5.1's input and output rates are twice those amounts, but its cache reads cost half as much.

A model in a higher price tier is therefore not necessarily more expensive for every kind of processing. Repeatedly reading long documents produces a different cost breakdown from generating large amounts of new text each time. Differences in token usage also change the cost of completing a task, even at identical rates.

Compare the total needed to reach the required quality, not only the unit price. A cheap first answer can become expensive after many revision requests. Conversely, a stronger model may reduce follow-up exchanges. This pricing table alone cannot establish which is always cheaper.

Finding vulnerabilities is allowed; exploit generation remains restricted

Fable 5.1 allows work that identifies software vulnerabilities, but retains restrictions that switch processing to Opus for activities such as exploit generation. In Claude Code, Anthropic expects cyber-related safeguards to intervene about 60% less often per session, on average, than with Fable 5.

The 60% figure is not a reduction in offensive capability or harm. It describes the developer's estimate for how often cyber-related safeguards interrupt work within a session. For users checking their own code, being able to finish a legitimate investigation matters alongside performance gains.

Enterprise customers can keep monitoring data under their own control

The accompanying Enterprise Frontier Safeguards (EFS) initiative stores abuse-monitoring data in a customer-controlled cloud. It automatically examines suspicious activity across conversations and allows the customer to handle any necessary human review.

Rollout is planned in stages from this autumn; it is not available to every company on the announcement date. Eligible customers will be offered zero-data-retention arrangements for Fable 5 and 5.1 until EFS is ready.

For organizations handling confidential information, greater model capability is not enough. Who stores activity records and who can read them also matters. This update can be read as an effort to bring frontier models into everyday business by addressing performance, ongoing costs and data management together.

Sources and references

• Introducing Claude Fable 5.1 and Claude Mythos 5.1 Anthropic / September 1 announcement / Current official page
• Claude API pricing Anthropic / Archived September 2, Japan time / Current official page
• Developing Enterprise Frontier Safeguards with our customers Anthropic / September 1 announcement / Current official page

Update history

• September 2, 2026Published.
• September 29, 2026Clarified that the roughly 60% decrease concerns cyber-related safeguard interventions per session.