On September 1, 2026, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. The two are the same underlying model shipped under different safeguard regimes: Fable 5.1 is generally available, while Mythos 5.1 is restricted to vetted organizations working in cybersecurity and the life sciences.
Three things changed at once, and they pull in different directions. Capability moved up, most sharply on long-running agentic and scientific tasks. Effective cost moved down, though not through the headline per-token rates. And Anthropic paired the release with a separate announcement, Enterprise Frontier Safeguards, that changes where regulated customers' activity data lives.
Anthropic
AWS
Google Cloud
MicrosoftWhat shipped
Fable 5.1 carries the API model ID "claude-fable-5-1" and is available across Claude.ai, the Claude API, Claude Code, Claude Enterprise, Claude Platform, and the three major cloud marketplaces. Effort is selectable across five levels, Low, Medium, High, XHigh and Max, with defaults set to High in Claude Code and Medium in Claude.ai and Claude Cowork.
That effort ladder matters more than it sounds, because Anthropic's central efficiency claim is framed around it: the company says Fable 5.1 reaches similar or better results than Fable 5 at low or medium effort, which is where most production traffic actually runs. A model that needs less thinking budget to hit the same quality bar is a cost story as much as a capability one.
Benchmarks: the gains are concentrated, not uniform
Anthropic's published launch numbers show a wide spread depending on the evaluation. The largest single jump is on Terminal-Bench-Science 0.1, an agentic science benchmark, where Fable 5.1 scores 52.6% against Fable 5's 24.7%, Claude Opus 5's 29.0% and GPT-5.6 Sol's 22.4%. On Zapier's AutomationBench the pattern repeats at smaller scale: 31.4% against 17.1% for Fable 5.
Where Fable 5.1 gains most over Fable 5
Launch-day scores published by Anthropic on 1 September 2026. Higher is better. Self-reported by the vendor, not independently audited.
The widest gap is on Terminal-Bench-Science, where Fable 5.1 more than doubles Fable 5. On CursorBench the four models sit within about six points of each other.
Elsewhere the deltas compress hard. On CursorBench 3.2.0 the four models sit inside a six-point band (73.4% for Fable 5.1, 70.5% for Fable 5, 70.0% for Opus 5, 67.2% for GPT-5.6 Sol). Humanity's Last Exam moves from 57.8% to 60.9% without tools, and from 63.8% to 65.0% with them. GDPval-AA v2, which scores economically valuable knowledge work on an Elo-style scale, goes from 1,723 to 1,853, with Opus 5 close behind at 1,824.
The consistent reading across those numbers: the generational gain sits in long-horizon, tool-using, multi-step work, and it is much smaller on single-turn reasoning. Mythos 5.1, the restricted twin, scores higher still on Terminal-Bench 4.0 at 60.9% against Fable 5.1's 55.8%, which is a safeguards difference rather than a different model.
All of these figures are vendor-published and self-reported. Independent replication on comparable harnesses usually lands lower than launch-day numbers, so treat them as a claim to be checked rather than a settled result.
Pricing: same list rates, a 75% cut on cache reads
Fable 5.1 keeps Fable 5's headline pricing exactly: $10 per million input tokens and $50 per million output tokens. The entire reduction lands on cache reads, the tokens a model re-reads from context it has already processed, which drop from $1.00 to $0.25 per million.
The Fable 5.1 price cut lands on cache reads, not on the headline rates
Change in cost relative to Fable 5, per Anthropic’s launch post. The workload figures are the vendor’s own estimates and vary with how much of a run hits cache.
Per-token list prices are identical to Fable 5. Whether a team sees 25% or 45% off depends almost entirely on cache-hit rate.
Anthropic estimates that works out to roughly 25% less for a typical workload and up to roughly 45% less for highly agentic work. The gap between those two numbers is the whole point. A long-running agent re-reads the same tool definitions, system prompt and accumulated transcript on every turn, so cache reads dominate its bill in a way they never do for a one-shot chat request. Teams whose traffic is short and varied will see very little of the advertised saving; teams running multi-hour autonomous sessions will see most of it.
Whether a team sees 25% or 45% off is not a negotiation. It is a measurement of their own cache-hit rate.
On the structure of the Fable 5.1 price change
That is a meaningful shift in how frontier pricing is being tuned. Rather than cutting the sticker price, which would apply evenly and cost the vendor across all traffic, the discount is aimed precisely at the usage pattern the vendor wants to grow.
Enterprise Frontier Safeguards: the data-residency answer
The second announcement is aimed at a specific blocker. Frontier-model safeguards generally require retaining some activity data for misuse detection, and regulated enterprises generally do not want that data sitting in a vendor's account. Enterprise Frontier Safeguards splits the two.
Under EFS, activity data is written into the customer's own cloud storage, Amazon S3, Azure Blob Storage or Google Cloud Storage, with the customer managing encryption keys, access policy and audit logging. Anthropic operates automated misuse detection across sessions and accounts, looking for patterns such as fraud, cyberattack preparation and credential theft, and routes alerts to the customer rather than to an Anthropic reviewer. Customer-owned storage, customer-managed keys and fully automated review are each optional, and none of them change model behavior, API pricing or rate limits.
Anthropic says it charges nothing for EFS itself; customers pay their cloud provider for storage, reads, writes and egress at standard rates. The program was built with input from more than 100 enterprises across financial services, healthcare, manufacturing, telecom, law, retail and the public sector, along with AWS, Google Cloud and Microsoft Azure.
Wells Fargo's chief information security officer described the arrangement in the announcement as keeping "custody of our data while Anthropic operates the detection," adding that "that split is what lets our teams put frontier models to work safely." Stripe's head of security said EFS "will enable us to use covered frontier models while retaining logs in Stripe's AWS environment."
Rollout is phased and begins in fall 2026, across Claude Code, Claude Enterprise, Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google Agent Platform and Microsoft Foundry. Until it lands, eligible customers can run Fable 5 and Fable 5.1 under zero data retention as an interim arrangement.
The design question worth watching is what happens when detection and review both sit on the customer's side. Alerts that once reached a vendor safety team now reach the customer's own security function, which is exactly what a bank wants and also a meaningful transfer of responsibility for acting on them.
Safeguards tuned for fewer false positives
Alongside the residency change, Anthropic reports retuning the classifiers themselves. Cybersecurity safeguards are described as firing roughly 60% less often than Fable 5's while still permitting vulnerability discovery and blocking exploit development. Biology safeguards are reported to trigger about 85% less often on benign elementary biology and medical questions, with research and development tasks steered toward Opus-class models instead.
The release also adds anti-distillation measures: new API accounts cannot manually edit prior context while preserving the transcript, closing a technique documented earlier in 2026 when competitors were accused of distilling Claude outputs. Anthropic additionally reports improved robustness on prompt-injection benchmarks, which is the failure mode that matters most as these models are handed browser and terminal access.
Every one of those figures is a vendor claim about its own systems. Reduced intervention rates are good news for users only if the reduction came from better precision rather than a lower threshold, and that distinction is not observable from outside.
The science results
The most striking claims in the launch material are not benchmark scores. Anthropic reports that the model designed protein binders with roughly ten times higher affinity than competition submissions, achieving close to a 50% hit rate across twelve targets where 10% to 15% is typical. It reports reprocessing 30-year-old radar data into a Venus elevation map at 2 to 3 kilometre resolution, against 10 to 20 kilometres previously. And it reports optimizing seven deep learning models by up to 2.5x through hand-written GPU kernels, cutting estimated costs 30% to 60% on genome-wide analyses.
These are anecdotes rather than evaluations, and they are selected by the vendor. They are still the more informative signal about where the model is being aimed: work measured in hours of autonomous execution against a verifiable result, not in single answers.
What this means for buyers
Fable 5.1 is the fifth Claude 5-generation model to ship since Mythos 5 in June, and the release pattern across the industry in 2026 is now clear. Labs ship families rather than flagships, and they differentiate inside a generation on price shape, effort ceiling and access tier rather than on one headline capability number.
For anyone choosing models, three practical points follow from this release. Cache-hit rate is now a first-order cost variable, not an implementation detail. Access tiering is real, and the strongest configuration of a model may not be the one you can buy. And the enterprise blocker has moved from capability to custody, which is what EFS is built to answer.
None of that makes any single model the right default. The benchmark spread in this release, wide gains on agentic science and near-parity on coding, is the same pattern that has held all year: leadership is distributed across evaluations, and the right model for a coding agent is often not the right model for document analysis or long-context research.
That is the case for keeping model choice portable. Claude Fable 5.1 is live on Metir from launch day, alongside GPT, Gemini and Grok models in the same workspace, so a task can be routed to whichever model actually performs best on it without a separate subscription for each new release.
Sources:
- Introducing Claude Fable 5.1 and Claude Mythos 5.1 - Anthropic
- Enterprise Frontier Safeguards - Anthropic
- Anthropic upgrades Claude with new Fable 5.1 model - 9to5Mac
- Claude Fable 5 and Claude Mythos 5 - Anthropic
- Claude Fable 5 API pricing and specifications - OpenRouter
- Introducing Claude Opus 5 - Anthropic
Image credits
Header image: Anthropic co-founder and chief science officer Jared Kaplan speaking at TechCrunch Sessions: AI, 2025, by TechCrunch via Wikimedia Commons, licensed under CC BY 2.0. In-body photo of Anthropic co-founder and CEO Dario Amodei at TechCrunch Disrupt 2023 by TechCrunch via Wikimedia Commons, licensed under CC BY 2.0.