Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. The two are the same model. What separates them is the safeguards each one runs: Fable 5.1 is generally available as claude-fable-5-1, while Mythos 5.1 goes only to vetted organizations through trusted access programs, with safeguards designed for cybersecurity and life sciences work.
That structure is the most interesting thing about this launch, because it produces a number the industry almost never gets to see.
Safeguards, priced in benchmark points
Anthropic published Terminal-Bench 4.0 scores for both configurations: 55.8% for Fable 5.1 and 60.9% for Mythos 5.1. Same weights, same benchmark, five points apart. The difference is the safeguards.
The company is unusually direct about the mechanism. Fable 5.1 was evaluated with its production safeguards switched on, and on tasks where those safeguards intervened, it scored zero. Anthropic's own footnote says this "likely reduces the performance of Fable 5.1 and Fable 5" on the affected benchmarks. Most labs would report the unrestricted number and move on. Publishing both, and flagging which one is depressed by its own safety layer, turns an abstract tradeoff into a measured quantity.
On the headline benchmarks, the generational jump is real. Terminal-Bench-Science 0.1, an agentic scientific research benchmark, goes to 52.6% from Fable 5's 24.7% — roughly double, with Opus 5 at 29.0% and GPT-5.6 Sol at 22.4%. AutomationBench nearly doubles as well, 31.4% against 17.1%. GDPval-AA v2 moves to 1853 from 1723, and Humanity's Last Exam to 60.9% without tools. Anthropic also notes that at Low and Medium effort, Fable 5.1 matches or beats Fable 5 at substantially lower cost, which matters more than the ceiling for anyone running these models in production.
Standard caveat: these are vendor-reported numbers on a vendor-chosen suite. The Terminal-Bench-Science figures carry a stated standard error of 3.5 to 4.5 points, and the OSWorld 2.0 scores use an August 2026 task release that isn't comparable to previously published results — which is why no competitor score appears in that row.
The price cut is a cache-read cut
Per-token pricing is unchanged at $10 and $50 per million tokens. What changed is cache reads, which drop 75% to $0.25 per million. Anthropic estimates this lowers real costs by around 25% for typical workloads and up to roughly 45% for highly agentic ones.
Read backwards, that's a disclosure about where the money in agentic work actually goes. If cutting a single line item by 75% takes 45% off the bill, then cache reads are most of the bill. Every long-running agent re-reads its accumulated context on every turn, so the cost of a task scales with how much history it drags along rather than with how much new text it writes. Anthropic's own chart makes the point explicitly: on a highly agentic workload, cache reads dominate the indexed cost bar, and "all other tokens" is the small remainder.
Enterprise Frontier Safeguards
The launch also addresses a constraint that has followed the Fable line since the export-control episode that saddled it with a 30-day retention regime: Fable-class models have not been available under zero data retention.
Enterprise Frontier Safeguards is Anthropic's answer. EFS stores data in cloud infrastructure the customer controls rather than Anthropic's, and any human review is done by the customer by default — privacy equivalent to a zero-retention agreement while still running misuse detection. Anthropic says it developed EFS with more than 100 customers and its cloud partners at AWS, Google Cloud, and Microsoft Azure, with a phased rollout starting this fall. Until it ships, eligible customers can use Fable 5.1 with zero data retention.
What the safeguards now permit
The safeguard changes are specific enough to be worth reading closely, because they draw a line rather than loosening a dial.
In cybersecurity, Claude Code users should see roughly 60% fewer safeguard interventions per session. The substantive change is that Fable 5.1 can now be used to discover software vulnerabilities, but not to develop exploits for them. Penetration testing, exploit generation, and binary-based vulnerability scanning still get redirected to Opus models. Find the bug, don't weaponize it.
In biology, safeguards now fire 85% less often on elementary biology and medical questions, a change that also applies retroactively to Fable 5. Life sciences research and development still routes to Opus models. Access to Mythos 5.1's advanced biology capabilities runs through a Life Sciences Verification Program that Anthropic developed with the US government, with enrollment for scientists expected to open soon.
The distillation countermeasure
Buried in the safeguards section is the change with the widest blast radius for developers. New API accounts can no longer manually edit Claude's prior context in a multi-turn conversation while preserving the transcript of its earlier thinking. That combination was a publicly documented technique for extracting a model's reasoning at scale, typically across thousands of fake accounts. At the API level, an edited history now returns a 400 on affected accounts.
Existing accounts aren't affected yet, but the restriction applies to everyone with future model releases, and a small number of custom integrations will break when it does. Anything that rewrites conversation history — compaction layers, harnesses that trim old turns, replay tooling — needs to become append-only.
The timing is not subtle. This is the first Anthropic flagship since the White House accused Moonshot of distilling Anthropic's Fable models, a fight in which Anthropic was cast as the victim. The launch materials still don't argue the policy question. They just close the hole.
A claimed system-prompt leak
On the same day, a researcher posting as "Pliny the Liberator" claimed to have extracted Fable 5.1's full system prompt, describing it as over 270,000 characters and posting a diff against their earlier capture of Opus 5's prompt. Anthropic hasn't commented. Claims like this are difficult to verify from the outside — the extraction techniques involved can produce plausible-sounding text without it being a genuine, verbatim copy of the model's actual instructions, so there's no way to confirm authenticity based on the claim alone.
The science results
Anthropic is making a broader claim with this release: that these models are close to contributing to scientific discovery rather than merely assisting with it. Three results are offered as evidence.
In molecular design, Mythos 5.1 designed protein binders that were validated experimentally by two external organizations. On three targets, binding affinities came in ten times higher than the best designs submitted to Adaptyv Bio's protein design competitions, and the hit rate reached nearly 50% across 12 targets — against the 10–15% that is typical in protein design today. That extends Claude's earlier work on protein design and chemistry from proposal into wet-lab confirmation.
In planetary science, Fable 5.1 trained a neural network on 30-year-old radar imagery from NASA's Magellan mission to build a new elevation map covering a third of Venus, resolving features at two to three kilometers instead of 10 to 20. Anthropic is releasing it under a Creative Commons license ahead of the upcoming NASA VERITAS and ESA EnVision missions.
In computational biology, Mythos 5.1 wrote custom GPU kernels that sped up seven open-source deep learning models by up to 2.5 times with identical outputs, cutting estimated GPU costs by 30 to 60% on large analyses. Anthropic frames this as days of work replacing weeks of specialist engineering that academic labs usually can't afford, and plans to open-source the optimizations. It lands alongside the Model Hardware Standard for operating lab equipment and the discounted Team plan for scientists, which together look less like scattered research demos and more like a deliberate push into the science vertical.
The EU compliance layer
There's also a regulatory first here. As a signatory to the EU AI Act's Code of Practice on Transparency of AI-Generated Content, Anthropic now watermarks the output of models released after August 2, 2026. The watermark is invisible without the detection API, contains no information about the user or their conversations, and Anthropic says it has no practical effect on output quality.
The Act also requires a way to check for it, so a detection API is rolling out in private preview to the categories EU law specifies — regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and civil society groups — plus enterprises with their own compliance obligations. It's a concrete instance of the pattern running through the EU's recent AI enforcement: obligations that arrive as shipped product features rather than published policies.