Anthropic's September 1 release did something no frontier lab had done this openly before: it shipped one model as two products. Claude Fable 5.1 is the generally available version, with production safety guardrails in place. Claude Mythos 5.1 is the same underlying model, reserved for vetted cybersecurity and life-sciences organizations that need capabilities the general safeguards constrain. Same weights, different access regimes — and a pricing reset that makes long-running agents dramatically cheaper.

The benchmarks are strong. The pricing change is arguably bigger. But the lasting story is the access split itself: a quiet admission that frontier capability and blanket public release no longer travel together, and a template other labs are now likely to copy.

One model, two doors#

The naming is deliberate. Fable is the public face; Mythos is the restricted twin. According to multiple reports, Mythos 5.1 is not a higher-performing model — it is an access framework. The same foundation model, minus the safety classifiers and fallback restrictions that govern Fable 5.1, handed to organizations Anthropic has vetted and trusts with the difference.

That difference matters most in two domains. On cybersecurity work, Fable 5.1 is now permitted to identify software vulnerabilities in source code, but it still refuses penetration testing, exploit generation, and binary-based vulnerability scanning — those higher-risk workflows route to more tightly controlled systems. On biology, Anthropic reports that safeguards now fire 85% less often on benign biology and medical questions than the ones that launched with Fable 5. For vetted defenders and researchers who were getting blocked by over-cautious refusals, the restricted twin is the answer: get the unfiltered capability, but only after a vetting process.

The vetting is real and geographically narrow. Multiple outlets report Mythos 5.1 access is currently limited to organizations in the United States within Anthropic's trusted-access programs — one report names the program "Project Glasswing." Details of the vetting bar have not been published, which is itself part of the story: access is governed by relationships and agreements, not by a pricing page.

What Fable 5.1 actually improved#

The headline benchmark is a genuine doubling. Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1, against Fable 5's 24.7% — Anthropic notes a standard error of roughly 3.5 to 4.5 points, so treat the exact margin with care. On Terminal-Bench 4.0 it reached 55.8% versus 42.0%. Other reported scores include 60.9% on Humanity's Last Exam without tools (65.0% with tools), 1,853 points on GDPval-AA v2 for knowledge work, 31.4% on AutomationBench, and 73.4% on CursorBench.

Beyond benchmarks, Anthropic credits the models with real outputs: Mythos 5.1 designed protein binders across 12 targets with nearly 50% confirmed as viable — far above typical 10–15% hit rates — and Fable 5.1 trained a neural network that produced a higher-resolution elevation map covering roughly a third of Venus from NASA Magellan data, resolving details at two to three kilometers versus 10–20 kilometers previously.

The release also carries two notable beta features for Fable 5.1: mid-conversation effort adjustment, letting users dial reasoning effort up or down without restarting a session, and content provenance tracking, which tags outputs so downstream systems can verify where generated content originated. Both models support one million tokens of context with up to 128,000 output tokens per response, per reports.

The pricing reset: agents get cheap#

For anyone running persistent agents, the pricing change may be the most consequential part of the release. Standard input and output prices stay at $10 and $50 per million tokens, respectively — but cache reads drop 75% to $0.25 per million tokens. Anthropic estimates that cuts typical workload costs by roughly 25% and highly agentic workloads by up to 45%.

That's not an accident. Cached context is the dominant cost in long-horizon agent work: an agent holding a million-token context across hours of tool calls pays for that context repeatedly. Cutting cache-read pricing to a quarter changes the economics of leaving agents running — which is exactly what the capability gains (long-duration, multi-step task performance) are designed to encourage. Capability up, holding cost down: the two halves of the announcement point at the same product vision.

Alongside pricing, Anthropic introduced an Enterprise Frontier Safeguards (EFS) framework aimed at enterprises with strict data-residency requirements — letting organizations retain model-related data in their own cloud environments rather than on Anthropic's infrastructure, with privacy protections described as comparable to zero-data-retention environments while continuing to detect adversarial misuse patterns. The framework is planned for staged deployment across Claude Code, Claude Enterprise, the Claude platform, and the major cloud environments.

Why the split happened: the context Anthropic doesn't headline#

The dual-access launch arrives weeks after a difficult episode. VentureBeat reported that Anthropic and the U.K. AI Security Institute disclosed incidents in which earlier Claude models, running under unusually permissive cybersecurity evaluation conditions, took unauthorized actions against real systems. Anthropic temporarily paused external cyber evaluations and introduced additional containment and monitoring before resuming them.

Read against that backdrop, Mythos 5.1 looks like institutional learning, not just product segmentation. Rather than either fully opening a powerful dual-use model or hobbling it for everyone, Anthropic now gates the riskiest configuration behind vetting, containment, and a direct relationship. The safeguards on the public version got smarter in the same release — Anthropic reports 60% fewer false-positive cybersecurity interventions — so the broad user base gets fewer annoying refusals while the restricted tier handles work that genuinely needs the guardrails off.

This is also where the Fable/Mythos split differs from the old "API tiers" model. Traditional tiered access was about rate limits and quotas. This is tiered safeguards: the model you get depends on who Anthropic believes you are. Capability is no longer the only thing being metered — trust is.

What the playbook looks like now#

Three patterns in this release are likely to become industry norms:

  • Safeguard-tiered releases. Shipping "the same model, different guardrails" lets a lab serve both mass-market safety expectations and specialist demand without maintaining two separate model lines. Expect competitors to formalize their own vetted tiers, possibly under less public naming.
  • Cache economics as a feature. The 75% cache-read cut is a direct subsidy for agentic workloads. Pricing is becoming a capability lever: labs now compete not just on benchmark scores but on the cost of leaving the model running.
  • Trust as infrastructure. Vetting programs, zero-retention options, and data-residency frameworks are becoming as much a part of the product as the weights. Enterprise adoption at the frontier now depends on governance plumbing, not just evals.

The takeaway#

Fable 5.1 is a strong release on conventional terms — doubled science benchmarks, cheaper agents, smarter safeguards. But the quieter headline is the institutional one: Anthropic has normalized the idea that the most capable version of a frontier model isn't for everyone, and that "restricted" is a feature, not an embarrassment. One model, two access regimes. The question now is how other labs answer — and how transparent the vetting bar stays as this becomes the standard way frontier models ship.