On June 9, 2026, Anthropic announced the simultaneous launch of two models: Claude Fable 5, described 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 is the first time a model of this class — a tier above Opus — has been made generally available.
The announcement is dense: a new model class, a novel safeguards architecture with automatic fallback, scientific results unusual for a launch post, and a government programme, Project Glasswing, stepping out of the shadows. This guide covers each part in order. If your question is more practical — what it costs, how to integrate it, what to test before June 22 — we have a separate article on adopting Claude Fable 5 in your business: pricing, API and use cases.
The Mythos class: a tier above Opus
Until now, the Claude line-up read simply: Haiku for speed, Sonnet for balance, Opus for demanding work. The June 9 launch makes a fourth tier official. The "Mythos" class designates models whose capabilities exceed what Anthropic previously considered reasonable to put in everyone's hands — the first member, Claude Mythos Preview, was deployed in April 2026 within a closed programme, via Project Glasswing, and never offered to the public.
Fable 5 changes that: it is a Mythos-class model, wrapped in safeguards that make it distributable to everyone. Anthropic is explicit about the hierarchy: Fable 5's capabilities "exceed those of any model we've ever made generally available", it is state-of-the-art on nearly all tested benchmarks, and the longer and more complex the task, the larger its lead over other models grows.
The distinction between the two models fits in one table:
| Claude Fable 5 | Claude Mythos 5 | |
|---|---|---|
| Underlying model | Identical | |
| Safeguards | Full (cyber, bio-chemistry, distillation) | Partially lifted depending on the partner's profile |
| Access | General public: API, Pro, Max, Team and Enterprise plans | Project Glasswing partners, then a trusted access programme |
| API pricing | $10 / M input tokens, $50 / M output tokens | |
The duo's very name sums up the strategy: Mythos is the raw model; Fable, its version made safe for the public. Anthropic frames this joint launch as a trade-off between two imperatives — bringing advanced AI capabilities "to as many users as possible, as quickly and as safely as we can" — and announces that even more capable models will arrive "in the coming months".
What Fable 5 can do: benchmarks and demonstrations
The common thread of the published evaluations is autonomy: Fable 5 and Mythos 5 "can work autonomously for longer than any previous Claude models". Four areas are documented.
Software engineering: the Stripe case
The most quoted result comes from early testing at Stripe: on a 50-million-line Ruby codebase, Fable 5 performed a codebase-wide migration in one day — estimated at over two months of manual work for a whole team. Stripe describes it as "months of engineering compressed into days".
The second result is more technical but at least as important for the bill: on FrontierCode, Cognition's evaluation measuring the ability to pass difficult coding tasks while meeting the standards of high-quality production codebases, Fable 5 scores highest among frontier models — even at medium effort, the intermediate reasoning-depth setting. Combined with better token efficiency than previous Claude models, this means that at higher quality, the model consumes less than you might fear.
Knowledge work: senior-level financial analysis
On Hebbia's Finance Benchmark, designed to evaluate senior-analyst-level reasoning, Fable 5 takes the highest score of any model, with marked gains in document-based reasoning, chart and table interpretation, and problem solving. Trading firm IMC reports that Fable 5 aced its trading-analysis evaluations nearly across the board: factual lookup, conceptual reasoning, root-cause analysis, expected-value analysis.
Vision: a new state of the art, with less scaffolding
Fable 5 becomes the reference model for visual tasks, with two capabilities put forward: extracting precise numbers from detailed scientific figures, and rebuilding a web app's source code from screenshots alone.
The most striking example is deliberately playful. Previous Claude models failed at Pokémon FireRed even with harnesses providing maps, navigation aids and game-state information. Fable 5 beat the game with a minimal, vision-only harness — raw screenshots, no extra information. The technical message behind the anecdote: the model needs far less software scaffolding to carry out long visual tasks, which reduces integration work accordingly.
Memory and long context: the agent that takes notes
Fable 5 "stays focused across millions of tokens" in long-running tasks and — this is the new part — improves its own outputs using its own notes. The chosen test bench is the deck-building game Slay the Spire: with persistent file-based memory, Fable 5's performance gain is three times larger than the one measured on Opus 4.8 under the same conditions, and the model reaches the game's final act three times more often. This is the capability that long-running agents depend on — the missing piece for systems that work on a case for hours or days without losing the thread or repeating their mistakes.
The announcement also ships demos that are less quantified but revealing of the autonomy level reached: a solar-system simulation with eclipse prediction derived from physics, autonomous Factorio sessions, VibeCAD — a 3D modelling tool with an embedded AI copilot — and a music-synchronised fluid simulation. One last note for API users: at the highest effort setting, Anthropic says Fable 5 reflects on and validates its own work before answering.
Mythos 5 and science: the other half of the announcement
The most unusual part of the post is not about code but biology. Anthropic documents what Mythos 5 — the safeguards-lifted version — produces in the hands of its teams and scientific partners.
- Drug design accelerated by a factor of ten. Anthropic's in-house protein design experts accelerated aspects of the drug design process by around ten times. In an internal study, Mythos 5, equipped with protein design and bioinformatics tools but with no human assistance, matches or beats skilled human operators: it chooses binding sites, selects and runs the tools, and recovers from its own failures. Nine of the 14 protein targets tested yielded strong candidates.
- Better than specialised models on their own turf. On an unpublished viral capsid assembly prediction task (adeno-associated viruses, candidates developed by Dyno Therapeutics), Mythos 5 outperforms dedicated protein language models — a generalist model beating the field's specialised tools.
- Scientific hypotheses judged better in blind comparison. Compared blind against Opus-class models, the molecular biology hypotheses formulated by Mythos 5 are preferred around 80% of the time; one of them has since been corroborated by an independent published study.
- A week of genomics, fully autonomous. Left for a week on a genomics project, the model assembled single-cell data covering 138 animal species and millions of cells, then trained a custom machine learning model that outperforms a recently published Science model — while being 100 times smaller. Anthropic says these results are intended for publication.
These results explain the launch's architecture: the same capabilities that formulate novel therapeutic hypotheses are the ones Fable 5's bio-chemistry classifiers lock away from the general public. Hence the two-model system.
The safeguards: how do you make a Mythos model "safe"?
This is the structural innovation of the launch, and the part that concerns every Fable 5 user. Rather than constraining the model itself, Anthropic surrounded it with classifiers: separate AI systems that analyse requests in three areas.
- Cybersecurity: vulnerability exploitation, offensive cyber tasks, agentic hacking.
- Biology and chemistry: deliberately broad coverage at this stage, including for example gene therapy design.
- Distillation: attempts to extract the model's capabilities to train competing models.
When a classifier triggers, the request is not refused: it is handled by Claude Opus 4.8, Anthropic's next-most-capable model, and the user is informed every time. The tuning is openly conservative: perfectly benign requests will sometimes trigger the fallback, but the total stays below 5% of sessions on average. For the other 95%, performance is identical to Mythos 5 — it is the same model.
On the robustness of these protections, the announcement offers three data points: an external bug bounty totalling over 1,000 hours of testing without a universal jailbreak being found; initial progress towards one jailbreak by the UK AI Security Institute within a brief testing window — disclosed transparently; and internal red-teaming concluding that jailbreak resistance is better than Opus 4.6. Anthropic describes this setup as broader than its previous "ASL-3" protections, which mainly targeted bioweapons-related queries. The full details are in the model's system card, which also includes an alignment assessment: measured misaligned behaviours (deception, cooperation with misuse) are judged low and similar to Opus 4.8.
The final layer: a mandatory 30-day retention on all Mythos-class traffic, including for business customers and on third-party platforms. The data is not used for training or for any non-safety purpose, all human access is logged, and deletion happens after 30 days in almost all cases. The stated goal is defensive: detecting novel and multi-request attacks, and reducing false positives. The practical compliance implications — processing records, legal basis, internal AI policy — are detailed in our business adoption guide.
Project Glasswing and trusted access: who gets Mythos 5?
Project Glasswing is the programme Anthropic runs in collaboration with the US government, through which Claude Mythos Preview has been deployed since April 2026 to cyberdefenders and critical infrastructure providers. Mythos 5 ships there as an upgrade: Anthropic describes it as the model with the strongest cybersecurity capabilities in the world, and cites critically important software already secured through the programme.
Access is set to widen along two axes. First, a systematic trusted access programme for cybersecurity organisations. Second — and this is new — an equivalent programme for biology: selected researchers, in fundamental as well as translational research, will receive a version of Fable 5 with the bio-chemistry safeguards lifted but the cyber protections kept in place. The first cohort is announced for the coming weeks. The overall logic is taking shape: one model, with safeguard perimeters modulated by how much trust each category of users has earned.
Pricing and availability: the essentials
Fable 5 has been available since June 9 on the Claude API (model ID claude-fable-5) and consumption-based Enterprise plans, at $10 per million input tokens and $50 per million output tokens — less than half the price of Mythos Preview, double Opus 4.8. Above all, it is included at no extra cost in Pro, Max, Team and seat-based Enterprise plans until June 22; from June 23, usage will go through credits, pending a return to subscriptions that Anthropic presents as the longer-term goal.
Two weeks of included access, then, to form your own opinion on real cases. The detailed cost calculations, the routing matrix between Fable 5, Opus 4.8, Sonnet 4.6 and Haiku 4.5, and the 5-step test plan are in the companion article dedicated to business adoption.
What this launch tells us about what comes next
Beyond the benchmarks, three signals are worth keeping. One: the cadence. Mythos Preview in April, Fable 5 and Mythos 5 in June, and "even more capable models in the coming months" — the pace is not slowing, and the gap between what exists under restricted access and what is public is now measured in months. Two: the distribution architecture. The triptych of classifiers, fallback to a less capable model and safety retention is probably the release blueprint for the next frontier systems; businesses are better off designing their integrations accordingly rather than treating it as a parenthesis. Three: science as the proving ground. When a launch post cites drug candidates and a genomics model headed for publication, the centre of gravity of use cases is shifting — from assistance tasks to entire projects carried out autonomously.
For a growing business, the practical conclusion is the one from our companion article: test on your own files before June 22, measure, and route each family of tasks to the right model. If you want help — model selection, API integration, fallback handling, compliance — tell us about your project: we get back to you within 48 hours with a concrete approach.
