Claude Fable 5 and Mythos 5: the complete guide to Anthropic's launch

On 9 June 2026, Anthropic announced two models at once: Claude Fable 5, presented as a Mythos-class model made safe for general use, and Claude Mythos 5, the same underlying model with some safeguards lifted, reserved for a small circle of cyberdefenders and critical infrastructure providers. It was the first time a model of this class, a tier above Opus, was made generally available.

This guide was first published on launch day and revised on 17 October 2026. Since June, a lot has happened: both models were suspended three days after launch, came back on 1 July, and were succeeded on 1 September by Fable 5.1 and Mythos 5.1. The next section sums that up; the rest of the guide explains what the launch contained, with every claim attributed to its source. For the practical side, cost, integration and routing, see Claude Fable 5 for business: pricing, API and use cases.

What happened after the launch

Date	Event	Source
9 June 2026	Fable 5 and Mythos 5 launched; Fable 5 on the Claude API as claude-fable-5	Anthropic
12 June 2026	Access to both models suspended for all users after US export controls on access by foreign nationals	Anthropic
30 June and 1 July 2026	Export controls lifted, then access restored, with new safeguards	Anthropic
1 September 2026	Claude Fable 5.1 and Mythos 5.1 released; Fable 5 becomes a legacy model, retirement not sooner than 9 June 2027	Claude documentation

The full story of the suspension is in our article on the 19 days offline and the redeployment.

The Mythos class: a tier above Opus

Until June, the Claude line-up read simply: Haiku for speed, Sonnet for balance, Opus for demanding work. The 9 June launch made a fourth tier official. The first Mythos-class model, Claude Mythos Preview, had been deployed in April through Project Glasswing, a programme Anthropic runs "in collaboration with the US government" for a limited group of cyber defenders and critical software infrastructure providers.

Fable 5 changed that. Anthropic wrote that its capabilities exceed those of any model it had made generally available, that it was state of the art on nearly all the benchmarks it tested, and that the longer and more complex the task, the larger its lead over Anthropic's other models.

	Claude Fable 5	Claude Mythos 5
Underlying model	Identical
Safeguards	Classifiers on cybersecurity, biology and chemistry, and distillation	Some safeguards lifted, depending on the partner
Access at launch	Claude API, consumption-based Enterprise, and Pro, Max, Team and seat-based Enterprise plans	Project Glasswing partners
Access on 17 October 2026	Legacy model, still available; Fable 5.1 is the current version	Organisations verified through Anthropic's verification programmes, such as the Cyber Verification Program
API price	10 dollars per million input tokens, 50 per million output tokens

The price was presented as less than half that of Mythos Preview. On Anthropic's price list, it is twice Claude Opus 4.8 (5 and 25 dollars).

What Fable 5 could do: the published results

The common thread of the published evaluations is autonomy: Anthropic wrote that the two models can work autonomously for longer than any previous Claude model. The announcement documents four areas. The benchmark tables in the announcement are images; the figures below are the ones stated in its text.

Software engineering: the Stripe case

The most quoted result comes from Stripe: in a 50-million-line Ruby codebase, Fable 5 performed a codebase-wide migration in one day, work that would have taken a team over two months by hand, in the words the announcement attributes to Stripe, "compressed months of engineering into days". On FrontierCode, Cognition's evaluation of difficult coding tasks held to production standards, Fable 5 scored highest among frontier models, even at medium effort. No score is given in the text.

Knowledge work

On Hebbia's Finance Benchmark, built to evaluate senior-analyst reasoning, Fable 5 took the highest score of any model. The trading firm IMC reported that it aced nearly all its trading-analysis evaluations. Again, no figures are published in the text.

Vision

Anthropic highlighted two capabilities: extracting precise numbers from detailed scientific figures, and rebuilding a web app's source code from screenshots alone. The playful example: previous Claude models failed at Pokémon FireRed even with harnesses providing maps and game state; Fable 5 beat the game with a minimal, vision-only harness.

Memory and long tasks

Fable 5 "stays focused across millions of tokens in long-running tasks", and improves its own outputs using notes in files. On the card game Slay the Spire, persistent file-based memory improved its performance three times more than it improved Opus 4.8's, and the model reached the final act three times more often. The announcement also shows demos: a solar-system simulation, autonomous Factorio sessions, a 3D-printable model designed in a browser CAD editor, and a fluid simulation. One customer quoted in it, Yusuke Kaji, notes that "at the highest effort, Claude Fable 5 reflects on and validates its own work".

Mythos 5 and science

The most unusual part of the post is about biology. Anthropic documents what Mythos 5 produced in the hands of its teams and partners:

• Drug design accelerated "around 10 times" on some aspects of the process, by Anthropic's in-house protein design experts. In an internal study, the model with protein design tools but no human help matched or beat skilled operators: 9 of 14 protein targets yielded strong candidates.
• Better than specialised models at predicting the assembly of AAV capsids, a component used to deliver gene therapies, on candidates developed by Dyno Therapeutics. No numbers are given.
• Hypotheses preferred about 80% of the time in blinded comparisons with Opus-class models on molecular biology; one has since been corroborated by an independent lab.
• A week of largely autonomous genomics work on single-cell data from millions of cells across 138 animal species, producing a model that outperformed one published in Science while being 100 times smaller.

These results explain the two-model architecture: the capabilities that propose new therapeutic hypotheses are the ones Fable 5's biology and chemistry classifiers keep from the general public.

The safeguards: how to make a Mythos model "safe"

Rather than constraining the model itself, Anthropic surrounded it with classifiers, separate systems that analyse requests in three areas: cybersecurity, biology and chemistry, and distillation (attempts to extract the model's capabilities to train other models).

At launch, a flagged request was not refused outright: it was answered by Claude Opus 4.8, and the user was told. Anthropic described the tuning as conservative: on average the classifiers triggered in less than 5% of sessions, and most biology and chemistry requests fell back at launch. On the API today, the documentation describes the mechanism differently: a declined request returns a normal response with stop_reason: "refusal", and developers can opt into a server-side retry on another model with the fallbacks parameter, in beta (Refusals and fallback).

On robustness, the announcement gives three data points: an external bug bounty of over 1,000 hours of testing without a universal jailbreak; progress toward one by the UK AI Security Institute within a brief initial testing window, disclosed; and partner testing with no harmful answers to single-turn requests on cyberattack planning, exploit development or defence evasion. The suspension three days later, after Amazon researchers found a bypass, showed the limits of those tests.

The last layer: a 30-day retention of all Mythos-class traffic, on Anthropic's surfaces and on third-party platforms. The data is not used for training or for any non-safety purpose, all human access is logged, and it is deleted after 30 days in almost all cases. With Fable 5.1, Anthropic announced Enterprise Frontier Safeguards, which keep customer data in customer-controlled storage, rolled out in phases from this autumn (Anthropic, September 2026).

Who gets Mythos-class access

In June, Mythos 5 went to Project Glasswing partners, and Anthropic announced a trusted access programme for cybersecurity organisations, then one for biology researchers. In September, with Mythos 5.1, the picture was: access only for a set of US organisations for now, a Cyber Verification Program and a Life Sciences Verification Program, and coordination with the US government to widen access to domestic and international partners (Anthropic).

Pricing and availability, then and now

At launch, Fable 5 was available on the Claude API and consumption-based Enterprise plans at 10 dollars per million input tokens and 50 per million output tokens, and included in Pro, Max, Team and seat-based Enterprise plans until 22 June, after which it required usage credits. After the redeployment, Pro, Max, Team and select Enterprise plans could use it up to 50% of weekly limits until 7 July, then through usage credits.

On 17 October 2026, Fable 5 is a legacy model. Fable 5.1 keeps the same base price, with cache reads at 0.25 dollar per million tokens instead of 1 dollar, and Anthropic now recommends starting most workloads on Claude Opus 5.5 (4 and 20 dollars), keeping Fable 5.1 for demanding reasoning and long-horizon agentic work (models overview).

What this launch told us about what comes next

What follows is our analysis.

Three signals hold up four months later. The cadence: Mythos Preview in April, Fable 5 in June, Fable 5.1 in September, so the gap between restricted and public capabilities is counted in months. The distribution architecture: classifiers, a fallback to a less capable model and safety retention have become the standard way these models ship, and integrations should be designed for refusals rather than surprised by them. And politics: a model can now be switched off by a government decision, not only by an outage. For anyone building on these models, the practical conclusion is the same: keep the model behind your own interface, so that switching to another one is a configuration change.