Anthropic said Mythos was too dangerous to release. Then it filed for an IPO.

That is not an Anthropic quote. It is, however, an accurate description of what Anthropic actually did on June 9, 2026, when it launched Claude Fable 5.1

Claude Fable 5 is a "public" version of Claude Mythos, the frontier-class model Anthropic spent two months telling the world was too dangerous to put in anyone's hands without extensive vetting.2 The company says it now beats every Claude model ever made generally available on nearly all benchmarks — software engineering, document reasoning, vision, long-horizon tasks.3

The benchmarks are real and the performance gains appear genuine. Stripe ran it against a 50-million-line Ruby codebase and reported a migration in one day that would have taken two months by hand. Hex said it's the first model to hit 90% on their core analytics benchmark. Anthropic also posted a video of Fable 5 beating Pokémon FireRed with nothing but raw screenshots — a party trick, but a technically impressive one.

None of that is the roast.

Here is what Fable 5 actually is: the same underlying model as Mythos 5, with a classifier layer bolted on top.3

When Fable detects that a query touches cybersecurity, biology, chemistry, or anything that looks like model distillation, it silently reroutes the request to Claude Opus 4.8 — the previous flagship, already a capable model, but not what you're paying for. Anthropic says this fallback triggers in fewer than 5% of sessions, and that users are told when it happens. The classifier is tuned conservatively on purpose, which means it will sometimes catch harmless requests. Anthropic's own words: it "will sometimes catch harmless requests."1

The pricing is $10 per million input tokens and $50 per million output tokens — double the price of Opus 4.8.3 When Fable silently routes your $50/Mtok query to an Opus 4.8 response, it is unclear what you're billed. That detail is absent from the launch materials.

Benchmark comparison of Fable 5 and Mythos 5 against competing models 3

For subscription users — Pro, Max, Team, and seat-based Enterprise — Fable 5 is included at no extra cost from today through June 22.1

On June 23, Anthropic removes it from those plans. After that, using Fable 5 requires usage credits. The company says it will restore Fable 5 as a standard subscription feature "as quickly as possible" and may extend the free window if capacity allows.

This rollout structure has a name: it's a free trial. Two weeks of flagship access to build enthusiasm, followed by a reversion to credit-based billing where heavy use generates variable monthly costs. Anthropic doesn't describe it as a free trial. The announcement frames it as a capacity management decision. Both things can be true simultaneously.

The launch announcement includes one paragraph that deserves significantly more attention than it has received.

Anthropic requires 30-day mandatory data retention on all traffic routed to Fable 5 and Mythos 5, on every platform — first-party and third-party.1 This applies even to enterprise customers who previously negotiated zero-retention agreements.

The stated reason is security: Anthropic wants to monitor for complex jailbreak attempts and reduce classifier false positives. The data will not be used for training and will be deleted after 30 days in most cases. Anthropic says it has added human access logging.

The framing is reasonable. The precedent is not.

Enterprise customers negotiate zero-retention agreements specifically because their queries contain sensitive business information — legal strategy, unreleased product details, M&A discussions, patient data in compliant deployments. The fact that Anthropic will not use this data for training is a different question from whether collecting it changes enterprise risk calculus. Anthropic is not asking whether customers want this trade-off. It is informing them they are making it.

If this policy holds for Fable 5 and Mythos 5, it establishes a norm: access to each successive tier of frontier capability comes with a new floor of mandatory data retention, framed as a safety necessity. The company said as much in its own announcement, noting the policy covers "Mythos-class models" going forward.1

Anthropic published its Fable 5 launch on June 9, 2026. One day earlier, the company filed confidentially for what could be the largest AI IPO in history.4

The same week, Anthropic's alignment research institute published a paper titled "When AI Builds Itself" — a formal warning that frontier AI systems are approaching recursive self-improvement, and that the industry should coordinate brakes on development.5

So in a seven-day window, Anthropic: warned that AI is becoming dangerously autonomous, filed for an IPO, and released the model they had previously described as too dangerous for general access.

Each of these actions is defensible in isolation. Taken together, they describe a company managing a specific tension: the brand value of being the "responsible AI" company depends on being seen as cautious, but the commercial case for an IPO at a high multiple depends on being seen as the capability leader. Fable 5 is the solution to both problems — released with enough safety theater to maintain the responsible brand, priced at double Opus 4.8 to signal that it's the real thing.

Claude Fable 5 is a genuinely powerful model wrapped in the most elaborate product release structure Anthropic has ever built.

The 13-day free window is a trial. The silent fallback to Opus 4.8 is an undisclosed downgrade on 5% of your Mythos-priced queries. The mandatory 30-day retention override is a unilateral change to enterprise data agreements, justified by safety language that enterprises cannot negotiate away. And the whole thing launched two days after the IPO filing, which is not a coincidence — it's a product announcement timed to a financial event.

Anthropic's safety credibility is real and has been earned over years of published research. This launch spends some of it. A company that genuinely believed its own warnings about Mythos-class models being dangerous would not release a version of that model with a free trial window and a 13-day subscription clock. Companies that believe their products are dangerous do not hand them out to drive IPO enthusiasm. They believe their products are useful, and the danger framing is the marketing.

The model is probably worth the money for engineering teams doing long-horizon agentic work. Read the data policy first.