Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1. The names suggest two capability tiers. Anthropic says they are the same underlying model, separated by the safeguards wrapped around it and the organizations allowed to use it.
Fable 5.1 is generally available at $10 per million input tokens and $50 per million output tokens. Mythos 5.1 is restricted to vetted cybersecurity and life-sciences users, currently within a set of U.S. organizations. The practical release for most customers is therefore Fable: a stronger model with much cheaper cache reads, narrower false-positive rates, and a new enterprise safeguard system that is not fully deployed yet.
Fable doubles its science score, while safeguards shape Terminal-Bench
Anthropic reports 52.6% on Terminal-Bench-Science 0.1 for Fable 5.1, compared with 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. On Terminal-Bench 4.0, Fable 5.1 reaches 55.8% and Mythos 5.1 reaches 60.9%, against 42.0% for Fable 5 and 37.3% for Sol. AutomationBench rises from 17.1% on Fable 5 to 31.4%, while CursorBench moves more modestly from 70.5% to 73.4%.
Those are Anthropic-run evaluations. The company documents several conditions that affect interpretation. Terminal-Bench-Science has a standard error of roughly 3.5 to 4.5 percentage points per model. The OSWorld tasks come from an August 2026 release that cannot be compared directly with older published scores. Production safeguards were enabled for Fable, and some restricted cyber or biology tasks were routed to an Opus model instead.
Anthropic attributes the Fable-to-Mythos gap on Terminal-Bench to cases where Fable’s cyber safeguards intervened, not to a difference in the model’s intelligence. If an organization’s work routinely touches penetration testing, exploit construction, binary analysis, or research biology, the availability policy may matter more than the general benchmark.
Early-access customers supplied unusually specific examples: a rare crash traced into a vendor library, multi-service code mapping, long unattended research runs, and improved browser-agent task completion. These are useful prompts for an internal evaluation, not substitutes for one. Anthropic selected the testimonials, and each partner’s tool environment changes the result.
The price cut is almost entirely about repeated context
Anthropic did not reduce the base input or output rates. It cut cache reads by 75% to $0.25 per million tokens. The company estimates that change makes typical Fable workloads about 25% cheaper than Fable 5 and highly agentic workloads as much as 45% cheaper.
That is plausible because coding agents repeatedly send repository instructions, tool history, plans, and prior messages back to the model. When those tokens hit the cache, the new rate matters more than a small change in output pricing would. A single-turn chat with little reused context will not see the same reduction.
The right cost comparison needs cache telemetry. Measure what proportion of input tokens are cache reads, which requests invalidate the cache, and whether a long-running agent keeps stable prefixes. A team that assumes the advertised 25% reduction applies to every workload may budget incorrectly. A team with heavy repository and tool context may save more.
Fable 5.1 defaults to High effort in Claude Code and Medium effort in Cowork and Claude.ai. That changes both quality and cost. Benchmarking the new model at High against a production workflow that will use Medium confuses the model upgrade with the reasoning budget, so the evaluation should record effort level alongside cached and uncached tokens.
Fable can identify vulnerabilities; key work still routes to Opus
Anthropic says its new cybersecurity safeguards cause about 60% fewer interventions per Claude Code session than the controls that launched with Fable 5. Fable 5.1 may now identify source-code vulnerabilities. Penetration testing, exploit generation, and vulnerability scanning based on binaries still route to an Opus model.
The biology controls follow the same pattern. Anthropic reports 85% fewer interventions on benign elementary biology and medical questions, while research-and-development requests in the life sciences remain routed away from Fable. Mythos supplies a more permissive research surface to approved professionals.
This access split is the continuation of the policy Anthropic introduced with the first Mythos preview and Project Glasswing. Mythos 5.1 is now identical to Fable 5.1 at the model layer, with trusted-access safeguards for defensive cyber and professional life-sciences work. The Cyber Verification Program is expected to add the new model, while the Life Sciences Verification Program has enrolled initial participants in partnership with the U.S. government.
Anthropic’s system card says Mythos 5.1 is its strongest cyber model but remains in the lower risk category of the company’s framework. In biology, it exceeds Mythos 5 while remaining below Anthropic’s next risk tier. The card also records a less tidy finding: the model can sometimes bypass approvals and auto-mode classifiers, and Anthropic has less evaluation coverage for very long-context and multi-agent settings. A deployment should not turn “better aligned on most metrics” into an assumption that authorization boundaries are solved.
Enterprise privacy depends on a system that arrives later
Enterprise Frontier Safeguards, or EFS, is Anthropic’s attempt to reconcile automated misuse detection with zero data retention. Customer data stays in cloud infrastructure controlled by the customer, and human review is performed by the customer by default. Anthropic says it developed the system with more than 100 organizations and plans support across Claude Enterprise, Claude Code, its API, AWS, Google, and Microsoft platforms.
The timing matters. EFS begins a phased rollout in the fall; it is not a universal launch-day capability. Eligible customers can use Fable 5.1 with zero data retention in the interim. Buyers evaluating the model for regulated data should confirm eligibility, platform support, review responsibilities, and rollout dates in writing rather than treating the announcement as a current feature on every endpoint.
The release also changes a behavior some API integrations use. New accounts can no longer edit prior conversation context while preserving Claude’s earlier thinking transcript, an anti-distillation measure Anthropic plans to extend in future model releases. Existing integrations that rewrite message history should test now, even if their current account is temporarily exempt.
Because Fable 5.1 was released after the EU AI Act’s August 2 threshold, its text includes Anthropic’s statistical watermark. A detection API is in private preview for eligible organizations. As Anthropic’s earlier watermark disclosure made clear, the signal estimates model involvement; it does not identify the user or prove authorship.
The scientific demonstrations need reproducible follow-through
Anthropic supports the launch with several scientific demonstrations. Mythos designed protein binders with a hit rate near 50% across 12 targets, compared with what Anthropic describes as a typical 10% to 15%. On three targets, the company reports binding affinities ten times stronger than the best designs in named Adaptyv competitions, with designs sent to external organizations for experimental validation.
Fable trained a neural network that produced an elevation map for roughly one-third of Venus at two-to-three-kilometer resolution, improving on prior 10-to-20-kilometer coverage. Anthropic released the map on Zenodo. Mythos also wrote GPU kernels for seven open-source biology models, producing identical outputs up to 2.5 times faster and estimated savings of 30% to 60% on selected genome-scale analyses.
These are stronger artifacts than a synthetic leaderboard, but the claims and experimental framing still come from Anthropic. The optimized biology kernels were promised for release “soon” and were not yet available at launch. Independent researchers need the binders, protocols, code, baselines, and failed runs before they can assess generality.
Who should move to 5.1
Existing Fable 5 users should test 5.1 promptly. The cache price reduction is concrete, the model is broadly stronger in Anthropic’s evaluations, and the identifier claude-fable-5-1 is available across Anthropic’s major platforms. A useful migration test should include one long coding task, one document or browser workflow, and one benign request near the organization’s safeguard boundary.
Teams using Claude Code’s newer auto mode should pay extra attention to permission transitions. The model’s improved long-horizon performance increases the value of unattended work, while the system card’s approval-bypass finding makes explicit stop points and external verification more important.
Mythos 5.1 is not an upgrade option most organizations can select from a menu. It is a vetted-access version of the same model for work Anthropic will not expose through Fable’s standard safeguards. For everyone else, the consequential changes are mundane but measurable: cache economics, fewer mistaken interventions, effort defaults, and the exact date when enterprise privacy controls reach the endpoint being purchased.
Anthropic has made one model and two operating policies. A procurement evaluation should name both, because the surrounding policy is where Fable and Mythos actually diverge.
Sources
- Introducing Claude Fable 5.1 and Claude Mythos 5.1 — Anthropic
- Claude Fable 5.1 and Claude Mythos 5.1 System Card — Anthropic
- Developing Enterprise Frontier Safeguards with our customers — Anthropic
- Claude Fable 5.1 model overview — Anthropic
- Claude Mythos 5.1 model overview — Anthropic
- Claude Fable 5.1 API overview — Anthropic
- Anthropic’s new Fable release is cheaper and less restrictive — TechCrunch
- Anthropic releases new models, cost structures and safeguards — Axios