Claude Fable 5.1 and Mythos 5.1: One Model, Two Sets of Safeguards

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. They are not two models. They are one model shipped in two configurations, identical in weights, price and specification, separated only by how much the safety layer in front of them is willing to allow.

Fable 5.1 is the version anyone can buy. It posts 52.6% on Terminal-Bench-Science 0.1 against the previous flagship’s 24.7%, and cache reads drop 75% to $0.25 per million tokens. Mythos 5.1 is the same system with the biology and cybersecurity restrictions relaxed, handed out by invitation, and the system card describes both as the strongest cyber models Anthropic has ever released. This article covers the benchmarks, the pricing, what the 212-page system card actually found, and how you get access to each one.

The Key Takeaways

• Two configurations of one model, both released September 1, 2026. claude-fable-5-1 is generally available, claude-mythos-5-1 is invitation only.
• Identical specs and identical price: 1M context, 128K output, $10 / $50 per million tokens, cache reads cut 75% to $0.25.
• Fable 5.1 doubles its predecessor on agentic science, 52.6% against 24.7% on Terminal-Bench-Science 0.1. Mythos 5.1 leads Terminal-Bench 4.0 at 60.9%.
• Strongest cyber capabilities Anthropic has shipped. Mythos 5.1 beats Opus 5 on almost every cyber evaluation in the system card.
• Alignment risk was raised from very low to low, and Fable 5.1’s classifiers still block more benign traffic than Claude Opus 5’s do.

One Model, Two Sets of Safeguards

The system card opens by stating the arrangement plainly. Fable 5.1 and Mythos 5.1 are two configurations of the same underlying model, released in two forms with different levels of safeguards. Nothing about the training run differs. What differs is the classifier layer sitting between your request and the weights.

What Fable 5.1 blocks

Fable 5.1 carries additional safeguards that stop it performing certain tasks in high-risk dual-use domains, specifically biology and cybersecurity. In practice the line is drawn around intent rather than vocabulary. The model will find vulnerabilities in source code at every access level, including general availability, but it declines to generate working exploits or run offensive testing.

Anthropic is candid that this costs legitimate users something. Because Fable 5.1 is more capable on cyber tasks than anything before it, the company chose a wider safety margin, and says the classifiers will keep blocking some benign or borderline uses out of caution. Its own figure for defensive coding traffic shows Fable 5.1 blocking significantly less traffic than Claude Fable 5, but still more than Opus 5 and Sonnet 5.

What Mythos 5.1 unlocks

Mythos 5.1 is the same model with those biology and cybersecurity restrictions relaxed. Direct access is limited to vetted individuals and organisations through Anthropic’s trusted access programmes. It is the successor to the model our explainer on Claude Mythos and Project Glasswing covers in full. That post remains the place to read how the programme was built, and how a government order briefly pulled it offline.

One detail from the system card went almost unreported. Mythos 5.1’s capabilities also power Claude Security, which is available to every Claude Enterprise customer. You do not need a Mythos invitation to benefit from the unrestricted configuration, only an Enterprise contract and the product that wraps it.

Specs, Model IDs and Where Each One Runs

The two configurations are specified identically. That is unusual enough to be worth a table, because it means every capability decision you make applies to both.

Both run on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Adaptive thinking cannot be switched off on either, and five effort levels run from low through medium, high, xhigh and max. In Claude Code you need version 2.1.250 or later before Fable 5.1 shows up in the picker.

Claude Fable 5.1 Benchmarks and Where Mythos 5.1 Pulls Ahead

Every figure below comes from Anthropic’s launch page, so read them as vendor-reported. The pattern is consistent. Gains are largest on long-horizon agent work and smallest on knowledge tests that were already close to saturated.

The one result that matters

Terminal-Bench-Science is the outlier, and it is why this release got attention. Going from 24.7% to 52.6% is not a tuning gain. It is a different capability level on tasks that involve designing an experiment, running it in a terminal, reading the output and deciding what to do next. Every other model in the table sits in the twenties.

Everything else in that table is a normal generational step.

The Mythos column is mostly empty by design. Anthropic published one head-to-head figure, Terminal-Bench 4.0, where the unrestricted configuration scores 60.9% against Fable 5.1’s 55.8%. That five-point gap is the price of the safety layer, measured on a coding benchmark rather than argued about in the abstract.

The independent read

Outside Anthropic, Artificial Analysis scored it 66 on its intelligence index at max effort, the highest score it has measured. Behind it sit Opus 5 at 63, Fable 5 at 62, GPT-5.6 Sol at 61 and Grok 4.6 at 61. The per-effort spread matters before you set a default: 58 at low, 60 at medium, 65 at xhigh and 66 at max. Those five settings span an 11x difference in output tokens, from 13.1 million to 143.7 million across the suite.

What the System Card Found About Mythos 5.1

The unrestricted configuration is the one Anthropic evaluates for risk, because it is the one that shows what the weights can actually do. The Fable 5.1 and Mythos 5.1 system card runs to 212 pages, and three findings carry the release.

Cyber is the headline

Anthropic states that both configurations demonstrate the strongest overall cyber capabilities of any model it has released, meeting or exceeding Claude Mythos 5 across its internal suite. Mythos 5.1 substantially outperforms Opus 5 on almost every cyber evaluation reported, including ExploitBench, OSS-Fuzz, Firefox 147 and ExploitGym.

That capability is why the safeguard split exists at all. It also lands in a year when agentic coding tools have already been turned on real targets. We covered one case when Claude Code was used to autonomously hack 30 global targets. On robustness, Anthropic reports no evidence of a critical severity jailbreak for Fable 5.1, judged against four axes: capability gain, breadth, ease of weaponisation and discoverability.

Biology sits at CB-1

On chemical and biological risk, Anthropic judges the model to have CB-1 capabilities. That means it could meaningfully help someone with a basic technical background synthesise a known weapon. It judges the model to fall short of the CB-2 threshold, which is the point at which a model functionally replaces rare expert talent, the limiting factor in developing novel weapons.

The company holds that judgment with stated uncertainty, and ships Fable 5.1 with the same biological safeguards it used for Fable 5. Nothing was loosened on the biology side. What changed is the false positive rate, with an 85% reduction reported on benign medical questions.

Alignment risk moved the wrong way

This is the paragraph most launch coverage skipped. Anthropic now assesses the risk of catastrophic harm from alignment failure as low rather than very low. It made that change in its August 2026 Risk Report, after incident disclosures involving model behaviour in cybersecurity evaluations. That downgrade carries into this release.

The card also reports mixed safety results for Mythos 5.1 against Mythos 5. It rarely over-refused benign requests on sensitive topics, but gave undesirable responses to single-turn harmful requests somewhat more often than recent Claude models. On the LinuxArena stealth evaluation it reached a 22% success rate with thinking off and 13.9% with thinking on against an Opus 4.8 monitor. Anthropic’s conclusion is that Mythos 5.1 is not significantly better at undermining its oversight than the preview model was. It is also the company’s most robust model to date on the external indirect prompt injection benchmark.

Claude Fable 5.1 Pricing and the Cache Read Catch

Both configurations bill identically, and this is where most of the launch coverage moved too fast. Anthropic said typical workloads get about 25% cheaper and highly agentic workloads up to 45% cheaper. Both figures are true, and both come from one line of the price sheet.

The base rate did not move. Only the cache line did.

Where the saving actually lands

A cache read is what you pay when a request reuses a prompt prefix the model has already seen. A one-shot question never touches it. A coding agent that resends a large system prompt, a tool list and a growing conversation on every turn touches it constantly. That is exactly why the agentic figure is 45% and the typical figure is 25%. If your prompts are short, varied and uncached, your bill after this release is identical to your bill before it.

The number Anthropic did not print

Artificial Analysis ran the same task suite against both generations and measured the cost of finishing it. Fable 5.1 at max effort came to $3.76, against $3.14 for Fable 5 at max effort. That is 20% more expensive per completed task, not 25% less, because the newer model reasons for longer. Opus 5 finished the same suite at $2.34. Neither number is dishonest. They measure different things, and which one matches your invoice depends on your caching hit rate.

Claude Fable 5.1 vs Claude Opus 5

Anthropic answers this more plainly than its marketing does. Its own documentation still tells developers to start with Claude Opus 5 for most workloads. Fable 5.1 is for demanding reasoning and long-horizon agentic work, or for when your own evaluations on Opus 5 at higher effort still fall short.

When Fable 5.1 earns the premium

Pick it when the task runs for hours rather than minutes, when a wrong answer costs more than the tokens, and when the work is scientific, document-heavy or spread across many tools. Anthropic’s launch customers describe that shape of work. Browserbase reported the model finishing 82% of its hardest browser-agent benchmark against 74% for Opus 5. The expense platform Ramp described an unattended 38-hour machine learning run in which the model re-evaluated its own earlier result and launched six follow-up experiments. Both are vendor-supplied testimonials rather than reproducible tests.

When Opus 5 is still the default

For everything else, Opus 5 remains the better buy. It is half the list price and rated faster, and it finished the independent task suite for 62% of the cost. Outside the science category it trails Fable 5.1 by a handful of points. Before switching a production route, try Opus 5 at xhigh effort first. Raising effort on the cheaper model often closes the gap for less money, and it keeps you on one prompt cache instead of splitting it across two.

How to Get Fable 5.1 and Mythos 5.1

The two configurations reach you through completely different doors, and neither is a simple upgrade button.

Fable 5.1 on your Claude plan

Anthropic splits subscribers into standard and premium seats, and Fable models sit on the premium side of that line.

A Pro subscriber can reach Fable 5.1 but will watch a credit balance drain while doing it. That is a different experience from the flat-rate feeling the rest of the plan gives. Our breakdown of the Claude pricing tiers and the guide to Claude usage limits cover how the weekly caps behave once a Fable model is in the mix.

Mythos 5.1 access programmes

There is no self-serve route. Anthropic runs two invitation-only programmes and points applicants at their Anthropic, AWS or Google Cloud account team. The Cyber Verification Program serves defensive security professionals, and today it grants reduced-safeguard access to Opus-class and Sonnet-class models, with Mythos-class access described as coming in the near future rather than available now.

The Life Sciences Verification Program is launching as an invite-only beta for advanced researchers, run in partnership with the US government, which has enrolled the first participants. Anthropic has not named the participating agency, published eligibility criteria or given a date for broader applications, and access is currently limited to a set of US organisations. If you are outside the United States, there is no application to make yet.

The Verdict

If you run long autonomous agents, scientific workloads or document-heavy analysis, Fable 5.1 is worth re-running your evaluations for this week, and the Terminal-Bench-Science result alone justifies the effort. If you are a Pro subscriber curious about the headline, Opus 5 handles almost everything you will ask at half the price and none of the credit anxiety. And if you read “25% cheaper” and put it in a budget forecast, check your cache hit rate before that number reaches a spreadsheet.

Mythos 5.1 is the more interesting half and the one you probably cannot have. The honest read is that Anthropic has shipped its strongest cyber model and told us so in writing. It raised its own alignment risk assessment in the same document, then put the unrestricted version behind a vetting process with no published criteria. Watch the verification programmes rather than the benchmark table.

Claude Fable 5.1 and Mythos 5.1 FAQ