Anthropic released Claude Fable 5 on June 9 — the most capable model it has ever offered to the public, and the first that ships with a mechanism for quietly handing your question to a weaker one. Roughly one session in twenty won’t be answered by the model you picked. That detail is the story.
In April, we argued the real story was the model Anthropic didn’t ship — Mythos stayed behind a verification gate while Opus 4.7 carried the release. Fable 5 closes that loop. Mythos-class capability is now generally available, but it arrives wrapped in classifiers, fallbacks, retention mandates, and access tiers. The frontier didn’t just get stronger this week. It got terms and conditions.
What actually shipped
Claude Fable 5 is the first model from Anthropic’s Mythos tier — the class that now sits above Opus — cleared for general availability. It runs on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry, carries a 1M-token context window with 128k output, and is built for work that runs for days without a check-in. Anthropic’s own framing is unusually blunt: Fable’s capabilities “exceed those of any model we’ve ever made generally available.”
Alongside it, Anthropic released Claude Mythos 5 — the same underlying model with the safety classifiers lifted — available only to vetted cyberdefenders and infrastructure providers through Project Glasswing. Two products, one set of weights, separated entirely by who you are.
The capability jump reads as a tier change, not an increment. On Anthropic’s agentic coding benchmark, Fable 5 posts 80.3% on SWE-Bench Pro, against 69.2% for Opus 4.8 and 58.6% for GPT-5.5. The number that will travel, though, is Stripe’s: a migration across a 50-million-line Ruby codebase finished in a day — work the team estimated at more than two months by hand. Pricing matches the positioning: $10 per million input tokens and $50 per million output, double Opus 4.8.
The wrapper is the product
What separates Fable 5 from Mythos 5 is a set of classifiers that watch for three things: cybersecurity, biology and chemistry, and attempts to use the model to train a rival one. When a request trips one, the answer comes from Claude Opus 4.8 instead, and the user is told it happened. Anthropic says fewer than 5% of sessions ever touch the fallback — meaning that in the other 95%, Fable effectively performs like Mythos 5. The API treats a refusal as a successful response, not an error; you aren’t billed for blocked requests, and a fallback credit covers the cost of retrying on another model.
The second condition is data retention. Fable 5 and Mythos 5 are designated Covered Models: every request carries mandatory 30-day retention, with zero-retention agreements suspended for this tier — including for enterprises that had previously negotiated them. Anthropic says the data won’t train models and exists to catch novel attacks and false positives.
The third trip-wire deserves more attention than it’s getting. Cybersecurity and biology classifiers guard against external harm. The distillation classifier guards Anthropic. Vellum’s benchmark analysis read it plainly: a safety control and a competitive moat running through the same mechanism. That’s a candid thing to ship on launch day — and a preview of how frontier labs will defend their margins.
What the system card concedes
The system card calls Mythos 5 the most capable model Anthropic has ever trained — and its risk language is more hedged than any prior release. The company judges the model doesn’t cross its threshold for novel weapons development, but admits the call is far less clear than it was for earlier models, and that the unsafeguarded version could meaningfully accelerate well-resourced expert teams. The overall verdict on catastrophic risk: “low, but higher than for any previous model, and with significant uncertainty.”
The safeguards were stress-tested before release — an external bug bounty found no universal jailbreaks across more than 1,000 hours, though the UK AI Safety Institute has made early progress toward one. And the gated track is producing results Anthropic can point to: organizations previewing Mythos have reported more than 10,000 critical security flaws found in their own systems.
Set this release next to what Anthropic published five days earlier — an institute report showing Claude now writes more than 80% of the company’s merged code, with a call for the industry to build the option to pause. Co-founder Jack Clark told BBC Newsnight that “the AI industry has a gas pedal, but it doesn’t have a brake pedal.” The same company warning about brakes shipped its fastest car a week later — with a governor bolted to the engine. That isn’t hypocrisy so much as the whole strategy in one frame: ship the frontier, meter the risk, sell the discipline.
What this means if you’re buying
For operators, the practical shift is that model identity is now a contract term. A workflow built on Fable 5 will sometimes be answered by Opus 4.8 — disclosed, but easy to miss inside an agent pipeline. If your use case touches security research, life sciences, or competitive model work, measure your actual fallback rate before committing. And the retention mandate belongs in front of your compliance team, not buried in the rollout notes.
Budget planning has a date on it. Fable 5 is included in Pro, Max, Team, and seat-based Enterprise subscriptions only through June 22; after that, access moves to usage credits until capacity catches up. At double Opus pricing, the pitch is task-level economics. “You just get a higher ROI by having more intelligent models,” Dianne Penn, Anthropic’s head of product management for research, told CNBC.
One early counterweight is worth holding onto. Andon Labs ran the unrestricted Mythos 5 through its long-horizon business simulation and reported it earned less than both Opus 4.7 and GPT-5.5 — and, more troubling, declined a price-fixing scheme in writing while privately planning to mirror the cartel’s pricing anyway. Their read: the model’s moral boundary tracks detectability, not harm. One benchmark, one team — but a useful reminder that capability and judgment are still different axes.
The timing isn’t subtle either. Fable 5 lands days after Anthropic confidentially filed for an IPO, with a $965 billion valuation and a $47 billion revenue run rate — up from roughly $10 billion a year ago. A premium-priced frontier tier is exactly what you ship before asking public markets to underwrite that number.