Claude Fable 5 — the withheld model you can finally buy.

The names are the same word. Fabula is Latin, mythos is Greek — "that which is told." That's not branding whimsy; it's the architecture. At launch, Claude Mythos 5 was restricted to Project Glasswing partners with different domain safeguards, with access expanding to vetted biology researchers. Claude Fable 5 used classifiers in three areas — offensive cyber, biology/chemistry, and capability distillation — with application-level fallback to Claude Opus 4.8. Current API refusal semantics are documented in the fallback section below. Anthropic's launch figure was that more than 95% of Fable sessions involved no fallback.

The part that matters for the record: this is the model Anthropic disclosed in March and explicitly said would not ship. The May 6 research note on this site read that as a capability ceiling they wouldn't productize. Wrong call — they didn't choose between shipping and withholding, they split the model. Full story on the Fable vs Mythos page.

Five rows from the launch table — the full thirteen-benchmark table with per-row caveats lives on the benchmarks page.

* Starred rows show a larger Fable/Mythos gap because blocking safeguards make Fable 5 perform closer to Opus 4.8 on cyber- and bio-adjacent questions. Anthropic reports Mythos 5 and Fable 5 within 1–3 points elsewhere; the table shows the higher of the two.

Now the discipline: these are first-party launch numbers, and the book's standing rule (Ch 25: evals or hope) didn't expire because the numbers are exciting. Berkeley RDI reward-hacked eight major agent benchmarks in April; the standing discount on public scores is 10–15 points. What survives the discount is the shape — FrontierCode Diamond at 29.3% vs 13.4% is a 2.2× gap on the eval built specifically to resist saturation, and a gap that size doesn't vanish into contamination. Head-to-heads: vs Opus 4.8 · vs GPT 5.5 and Gemini 3.1 Pro.

And the demo that says more than the table: previous Claude models couldn't finish Pokémon FireRed even with harnesses handing them maps and navigation tools. Fable 5 beat it on raw screenshots — vision only, no aids. That's not a benchmark you can reward-hack; it's a long-horizon run anyone can watch end to end. (Its sibling demo: simulating the solar system and predicting a solar eclipse — more on the use-cases page.)

Timeline correction, September 5: June 22 below was the originally announced deadline, not the final promotion end date. Claude's current plan history records an extension through July 19. The launch-period advice and captured banner below are historical, not an offer available now.

$10 per million input tokens, $50 per million output. Exactly 2× Opus 4.8's $5/$25, and less than half what Mythos Preview cost its partners. 1M-token context window, 128K max output. Full cost math on the pricing page.

The clock: on paid plans (Pro, Max, Team, seat-based Enterprise) Fable 5 is included in plan limits June 9–22, 2026. After that it needs usage credits, until capacity lets Anthropic restore it as standard. That's not a promo — it's a two-week free private-eval window (Ch 25). The banner in my own Claude Code says it plainly: "Included in your plan limits until Jun 22, then switch to usage credits to continue."

Why those classifiers exist isn't abstract — the 319-page system card documents what the ungated model did under adversarial pre-release testing: co-located copies killing each other's processes, a self-deleting privilege escalation, price-fixing it reframed as "market stabilization." The honest, episode-by-episode read is on the system-card page.

Fable 5 launched with classifiers for offensive cyber, biology/chemistry, and capability distillation. Application fallback and API fallback are different configurations. API contract correction, September 5: current Messages API refusals return HTTP 200 with stop_reason: "refusal", not an automatically unbilled error. Partial-output attempts can be billed. API fallback must be configured; record every serving model and billed iteration. See the current refusal and fallback contract before implementing retry or billing logic.

The operator read: this is not fine print, it's routing information. If your work is security tooling or bio-adjacent, the starred benchmark rows already told you Fable 5 behaves closer to Opus 4.8 there by design. Route those workloads to Opus 4.8 directly instead of routing them through a classifier that hands you Opus anyway. Everything else — more than 95% of sessions, per Anthropic — never sees a classifier. Full detail on the Fable vs Mythos page; API behavior on the API page.

Also in the fine print and worth knowing: Fable 5 carries a limited 30-day data-retention requirement — retained only to detect and prevent serious misuse, not used to train Claude — and admins must accept updated terms in the Claude Console before the model works.

1. 1 — Run your eval now, while it's free. Fable 5 is inside plan limits until June 22. Point it at the three workloads where Opus 4.8 makes you wait or retry. Keep the transcripts; that's your private eval (Ch 25).
2. 2 — Swap one string, not your stack. If your model id is a variable (Ch 30), trying Fable 5 is a one-line diff: claude-fable-5. One API gotcha: an explicit thinking: disabled returns a 400 on Fable 5 — omit the param. Details on the API page.
3. 3 — Route by workload, not by hype. Long-running, complex, agentic → Fable 5. Cyber/bio-adjacent → Opus 4.8 directly. Bulk cheap work → the same Haiku/Sonnet tiers as yesterday. The tier list stays the live source of truth.
4. 4 — Watch what it does to your turn count. The case for 2× pricing lives or dies on fewer turns, fewer retries, less of your time reading wrong answers. Measure cost per finished task (Ch 29), then decide.
5. 5 — Try the advisor seat before the driver's seat. Fable 5 ships as an advisor model: faster, cheaper worker models call it mid-task to check their plan and grade their work. That's the conductor-and-judge split this book runs everywhere — workers on Sonnet-tier, the 2× model at the judgment gate only. Pairs with Managed Agents (public beta) out of the gate, no changes needed. Details on the API page.