Anthropic Claude Fable 5: The First Public Mythos-Class AI Model Explained

"Fable 5's capabilities exceed those of any model we've ever made generally available."

— Anthropic, Claude Fable 5 announcement, 9 June 2026

On June 9, 2026, Anthropic released Claude Fable 5, the first publicly available version of its most powerful model family, internally known as Mythos. This is the same underlying technology that, after a limited partner-only preview in April, sent a jolt through the cybersecurity community for its ability to find and exploit vulnerabilities. What makes Fable 5 notable is not just raw capability — it is that Anthropic decided the only responsible way to ship a Mythos-class model to the public was to wrap it in hard safety guardrails. For enterprise leaders weighing where AI is genuinely headed, Fable 5 is one of the most consequential releases of the year. Here is what it actually does, what it costs, and where the caveats are.

Fable 5 and Mythos 5: What Is the Difference?

Anthropic released two models on the same day, and the distinction matters. Claude Mythos 5 is the full, unrestricted model. Claude Fable 5 is that same model with safety classifiers layered on top to make it safe for general availability. In other words, Fable is not a smaller or weaker model — it is Mythos with guardrails. Mythos 5 itself remains restricted: cybersecurity safeguards are only lifted for vetted partners in Anthropic's Project Glasswing program, and a separate biology trusted-access program is planned for select researchers.

This two-track approach is the headline. Anthropic is effectively saying that its frontier model is too capable in certain domains to release without restriction, and that the public version exists only because the company built classifiers it trusts to contain those capabilities. As Anthropic put it bluntly: "Without safeguards, Fable 5's capabilities in areas like cybersecurity could be misused to cause serious damage."

It Can Work Autonomously for Days

The capability that most distinguishes Fable 5 from previous Claude models is endurance on long-horizon work. According to Anthropic and to AWS, the model "can handle complex tasks that previous models could not sustain, running for days on coding and knowledge work without intervention." When run inside an agent harness, it plans across stages, delegates to sub-agents, writes its own tests to verify its work, and uses vision to critique its own output against the goal.

This is the practical embodiment of the agentic loop that has been the central promise of 2026's AI roadmaps: an agent that plans, acts, checks, and corrects without a human in the loop on every step. It is also why Fable 5 belongs in the same conversation as the rise of autonomous AI agents in the enterprise. Anthropic was direct about where the advantage concentrates: "The longer and more complex the task, the larger Fable 5's lead over our other models."

The Benchmark Numbers

Anthropic reports that Fable 5 is state-of-the-art on nearly all benchmarks it tested, and on some it scored more than 10% higher than Claude Opus 4.8, the previous flagship. Several concrete data points stand out:

• Coding (Stripe): Fable 5 compressed months of engineering work into days on a 50-million-line Ruby migration.
• Analytics (Hex): it was the first model to score 90% on Hex's core benchmark of complex, long-running analytical tasks.
• Finance (Hebbia): it posted the highest score on Hebbia's Finance Benchmark for senior-level reasoning.
• Memory: when given persistent file-based memory, its performance improved three times more than Opus 4.8 did under the same conditions.

The finance result is particularly relevant for the kind of document-heavy, senior-reasoning work that defines enterprise back-office functions. It is a useful signal for anyone tracking how AI is reshaping finance, where the bottleneck has always been reliable reasoning over messy, high-stakes data rather than raw text generation.

How the Safety Guardrails Actually Work

Fable 5 ships with three classifier systems that monitor its activity and, when triggered, hand the task off to the safer Claude Opus 4.8 instead. The three high-risk domains are cybersecurity (offensive and exploitation tasks), biology and chemistry (bioweapons and dangerous biological research), and distillation (attempts to extract the model's capabilities to train a competitor).

The numbers Anthropic published suggest the guardrails are designed to be unobtrusive for normal use: more than 95% of Fable sessions run entirely on the model's own responses, with the fallback to Opus 4.8 triggering in fewer than 5% of sessions on average. Anthropic also ran an external bug bounty that, in over 1,000 hours of red-teaming, "produced no universal jailbreaks," and the company says Fable 5 complied with zero harmful single-turn cybersecurity requests across 30 public jailbreak techniques. One trade-off enterprises should note: Anthropic imposes a mandatory 30-day data retention policy on all Mythos-class traffic, used to defend against novel attacks. The company states this data is not used for model training or non-safety purposes.

What It Costs, and When the Free Window Closes

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens — roughly double the cost of Opus 4.8, though Anthropic notes it is less than half the price of the earlier Mythos Preview. The existing 90% prompt-caching discount on input tokens still applies, which materially changes the economics for agentic workloads that reuse large context windows.

There is a timing detail worth flagging for anyone on a subscription. From June 9 through June 22, 2026, Fable 5 is included at no extra cost in Pro, Max, Team, and seat-based Enterprise plans. On June 23, Anthropic pulls it from those plans and moves to a credit-based model, meaning usage will then draw down purchased credits. If you want to evaluate the model against your own workloads without incremental cost, that two-week window is the moment to do it.

Where You Can Run It

Fable 5 is broadly available from day one: on the Claude Platform natively, on Amazon Web Services (both through Amazon Bedrock and the Claude Platform on AWS), on Google Cloud, and on Microsoft Foundry. For enterprises, the Bedrock path is notable because it runs within existing AWS environments with enterprise security and regional data residency, and it integrates with Bedrock's Advanced Prompt Optimization, which benchmarks prompts against evaluation criteria and outputs production-ready rewrites.

The model's strength on documents, diagrams, charts, and tables makes it a natural fit for document-heavy industries such as finance, legal, and analytics — exactly the functions at the center of most enterprise transformation programs. It is one more data point in the broader argument that AI is moving from interface to infrastructure inside the enterprise stack.

The Bottom Line

Claude Fable 5 is significant for two reasons that pull in opposite directions. On capability, it is the strongest model Anthropic has ever released publicly, with a decisive lead on long, autonomous, high-complexity work and standout results in coding, analytics, finance, and scientific research. On safety, its very existence is an admission that frontier capability now requires hard containment — the public only gets Fable because Mythos itself is considered too dangerous to release unrestricted.

For enterprise leaders, the pragmatic move is the same as with any frontier model: pilot it on a genuinely hard, long-running task where its endurance advantage can show up, measure the result against a cheaper model, and weigh the governance implications of the retention policy and the premium price. Fable 5 is not the model you reach for on every prompt. It is the one you reach for when the task is too big, too long, or too complex for anything else — and that is a category that is only going to grow.