2026-06-09 — views

Anthropic ships Claude Fable 5: a Mythos-class frontier model the public can finally touch

Read this because One frontier model, two releases: a safeguarded public version and a restricted full version. The pricing, the free window, and the safety-classifier fallback are what builders should check first.

Anthropic released Claude Fable 5 — its first publicly available Mythos-class model — at $10/$50 per 1M tokens, with a restricted Mythos 5 for vetted partners.

What shipped

On June 9, Anthropic released two models that are, underneath, the same thing. Claude Fable 5 is the company’s new flagship — a “Mythos-class” frontier model wrapped in hard safety classifiers and made available to everyone. Claude Mythos 5 is the same underlying model with the cybersecurity safeguards removed, restricted to pre-approved organizations: Project Glasswing cybersecurity partners today, with trusted-access programs for biology researchers and a broader cybersecurity tier planned.

The naming is the announcement’s quiet thesis. Fable comes from the Latin fabula, a sibling of the Greek mythos — Anthropic says the names distinguish the models “primarily by safeguards rather than underlying capabilities.” This is the first time a Mythos-class model has been publicly accessible at all; until now that tier lived behind a limited preview.

The numbers Anthropic put on the board

Item	Figure
API pricing	$10 per 1M input tokens / $50 per 1M output tokens
vs. Mythos Preview pricing	less than half
Sessions running fully on Fable 5	roughly 95% (fallback to Opus 4.8 in about 5%)
Jailbreak bug bounty	1,000+ hours, no universal jailbreaks found
Stripe case study	50M-line Ruby codebase migration in one day (vs. ~2 months manually)
Protein design case study	~10x acceleration; strong candidates for 9 of 14 targets
Subscription availability	included on Pro/Max/Team/Enterprise June 9–22, usage credits after

The pricing is the headline for builders. At $10/$50 per million tokens, Fable 5 costs double Opus 4.8 but lands at less than half of what Mythos Preview charged. Anthropic is clearly pricing this to be used in anger, not sampled.

How the safety split actually works

Fable 5 is not a smaller or dumber model — it is Mythos 5 plus AI-based classifiers that intercept three categories of requests: offensive cybersecurity, dual-use biology and chemistry, and distillation attempts (extracting capabilities to train another model). When a classifier fires, the session silently defers to Claude Opus 4.8 instead of refusing outright. Anthropic says this fallback triggers in under 5% of sessions, concentrated in security-adjacent work.

External red-team organizations and a 1,000-plus-hour bug bounty failed to find universal jailbreaks, per the announcement, and external testing showed zero compliance with harmful single-turn cyberattack requests. The trade-off: all Mythos-class traffic — including Fable 5 — carries a mandatory 30-day retention window, which Anthropic says is used only for safety and defense, never training, with logged human access and deletion after 30 days in nearly all cases.

Why this is a builder story

Three things matter if you ship software. First, the capability jump is aimed at long-horizon work. Cursor’s CEO called it “the state of the art model on CursorBench” and said it “opened up a class of long-horizon problems that were out of reach.” GitHub’s CPO echoed the autonomy claim. Stripe’s one-day, 50-million-line Ruby migration is the kind of number that, even discounted heavily, changes how you scope refactoring projects.

Second, the free window is real but short. Subscription users on Pro, Max, Team, and seat-based Enterprise get Fable 5 included from June 9 to June 22; after that it requires usage credits until Anthropic restores capacity. API and consumption-based Enterprise access is live immediately. If you want to evaluate it on your own workload, the next two weeks are the cheap window.

Third, the classifier fallback is a behavior change you can hit silently. If your product touches security tooling, vulnerability triage, or bio-adjacent research, some fraction of sessions will quietly run on Opus 4.8 instead. That’s not a refusal you can catch in logs by string-matching — it’s a different model answering.

The catch worth naming

The early benchmark quotes are all from launch partners, and launch-partner numbers always flatter. The 95%/5% classifier split is an aggregate; for a security-focused team the fallback rate could be far higher, making Fable 5 effectively Opus 4.8 with extra steps for exactly the work you bought it for. And the 30-day traffic retention — even with Anthropic’s no-training pledge — will be a compliance conversation in regulated shops before any production rollout.

Practitioner note

What I would do this week: run a fixed internal eval — ten representative long-horizon tickets from my own repo — on Fable 5 during the free window, scored on pass rate, turns to completion, and token spend versus Opus 4.8. At $50 per million output tokens, the “fewer turns” claim matters more than raw quality: a model that finishes in three turns can beat a cheaper one that takes nine. I’d also instrument for the fallback: log which model actually answered each session if the API surfaces it, and flag any security-related workflows for manual review. I would not build anything load-bearing on the June 9–22 included tier — treat it as evaluation budget, then decide with real numbers.

Under-considered angle

Everyone is reading this as a model release. The more durable shift is the two-tier release pattern itself: one trained artifact, two products separated only by safeguards and access policy. If this sticks, “frontier access” becomes a vetting and compliance product, not a capability product — and the interesting competition moves to who controls the trusted-access lists. Watch the distillation classifier especially: it means the public version of a frontier model now actively resists being used as a teacher for someone else’s training run. That quietly raises the cost of fast-following for every lab that built its strategy on distilling the leader.