Anthropic ships Claude Fable 5.1 and cuts the price of cache reads

Anthropic released Claude Fable 5.1 on September 1, 2026, together with Claude Mythos 5.1. The two are the same model under two safeguard configurations: Fable 5.1 goes to everyone, and Mythos 5.1 only to vetted organizations through two access programs. For a developer the release is less about a new name than about three numbers that change what a long agent run costs and what the model agrees to do.

What the price change actually is

Input and output prices did not move. Fable 5.1 bills 10 dollars per million input tokens and 50 dollars per million output tokens, the same as Fable 5. The cache read is what changed. On Fable 5.1 and Mythos 5.1 a cache hit costs 2.5% of the base input price, which works out to $0.25 per million tokens; every earlier model uses the standard 10% multiplier, so Fable 5 charged $1 for the same read.

Anthropic presents the saving as an estimate rather than a measurement, and that distinction carries weight. The company says workloads billed by token will cost about 25% less on average, and that heavily agentic work will often save considerably more, up to roughly 45%. Both figures depend on how much of a request is served from cache, which is a property of the workload rather than of the model. An agent that replays a large system prompt and a long tool history on every turn sits at the top of that range. A single call with no cache at all sees no change.

Where the safeguards moved

Fable 5.1 is now permitted to find software vulnerabilities. Developing exploits from them is still refused, and Anthropic says penetration testing, exploit generation and binary-based vulnerability scanning continue to route to its Opus models. The company expects Claude Code users to see around 60% fewer interventions per session from the cyber safeguards than they saw on Fable 5.

Three further changes ship in the same release, and each of them lands on a different group of customers.

• Enterprise Frontier Safeguards keep customer data on the customer's own cloud infrastructure instead of Anthropic's, and roll out in phases starting this fall.
• Until that system is ready, eligible customers may run Fable 5.1 under a zero data retention arrangement.
• The biology safeguards fire 85% less often on benign elementary biology and medical questions than the ones that launched with Fable 5.
• API accounts created from the release date onward can no longer edit Claude's earlier context in a multi-turn conversation while preserving the transcript of its earlier thinking, which closes a documented distillation route.

The benchmark numbers, and what carries them

On Terminal-Bench-Science 0.1 Anthropic reports 52.6% for Fable 5.1, against 24.7% for Fable 5, 29.0% for Claude Opus 5 and 22.4% for GPT-5.6 Sol. On Terminal-Bench 4.0 the figures are 55.8% for Fable 5.1 and 60.9% for Mythos 5.1, against 42.0% for Fable 5. On the GDPval-AA v2 knowledge-work score Fable 5.1 reaches 1853 against 1723.

These are the vendor's own runs, and the announcement is unusually direct about their limits. Anthropic states a standard error of 3.5 to 4.5 points per model on the science benchmark, and reports that its harness reproduces the public leaderboard scores for Opus 5 and Fable 5 at slightly different values, both within noise. It also states that Fable 5.1 was evaluated with production safeguards enabled and scored zero wherever those safeguards intervened, which pushes its own figures down rather than up.

Where you can run it

Fable 5.1 is available on the Claude API and, according to Anthropic, on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Azure. Amazon confirmed the same day that the model is generally available on Amazon Bedrock and on Claude Platform on AWS.

GitHub made it generally available in Copilot on the Pro+, Max, Business and Enterprise plans, across Visual Studio Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, the Copilot CLI, the coding agent and github.com. Two conditions come with that. Administrators on Business and Enterprise have to switch the policy on, because it is off by default, and GitHub states that the model requires data retention by default for Anthropic's safety classifiers.

What the science results are, and are not

The announcement leads with research rather than with coding. Anthropic says Mythos 5.1 designed protein binders with a hit rate close to 50% across 12 targets, where 10% to 15% is the norm in the field, and that on three targets the binding affinities came out 10 times higher than the best entries in Adaptyv Bio's design competitions. It also says the model wrote GPU kernels that sped up seven open-source deep learning models by up to 2.5 times with identical outputs.

Part of this was checked outside the company: the binder designs went to two external organizations for experimental validation. The rest are internal results. The elevation map of Venus that Claude built from Magellan radar data is being released under a Creative Commons license, which at least makes that one output inspectable.