Claude Fable 5: Anthropic's Mythos-class model goes public

Anthropic released Claude Fable 5 on June 9, 2026 — the first Mythos-class model anyone can actually use, and the first model in the Claude 5 line. It tops Anthropic's internal SWE-Bench Pro testing at 80.3%, eleven points ahead of Claude Opus 4.8 (69.2%) and more than twenty ahead of GPT-5.5 (58.6%), and it costs $10 / $50 per million tokens — exactly double Opus 4.8. One framing you will see everywhere today is wrong in an interesting way: Fable 5 is not a scaled-down Mythos. Per Anthropic's announcement, it is the same underlying model as the restricted Claude Mythos 5 — what differs is the safeguard configuration. High-risk queries (offensive cyber, biology, chemistry, distillation attempts) get blocked, and the answer is generated by Opus 4.8 instead. That fallback triggers in under 5% of sessions; the other 95%+ run on full Mythos-class weights.

How does Fable 5 rank in June 2026?

The numbers below are Anthropic-published, so read them as marketing-with-receipts — but the gaps are large enough to survive a generous error bar. The pattern from the launch material: the longer and more complex the task, the larger the lead. For shorter tasks the gap to Opus 4.8 narrows considerably.

Beyond the headline pair, Anthropic reports the model scored highest among frontier models on Cognition's FrontierCode benchmark even at medium effort, was the first model to clear 90% on Hex's benchmark of long-running analytical tasks, and topped the Hebbia Finance benchmark. The launch anecdote that matters most for developers comes from Stripe: a codebase-wide migration across a 50-million-line Ruby codebase, finished in one day — work Stripe estimated at more than two months for a team. It also completed Pokémon FireRed with a minimal agent harness, which earlier Claude models could not manage even with extra tooling. For where the rest of the field sits, see the full coding-model comparison.

What is Claude Fable 5?

Mythos-class is a new tier in Anthropic's lineup, sitting above Opus the way Opus sits above Sonnet. The first model in that tier, Claude Mythos Preview, shipped in April 2026 to a closed group of cyber defenders and critical-infrastructure providers under Project Glasswing, run in collaboration with the US government. Inside that program it found thousands of critical and severe vulnerabilities across major operating systems and browsers — which is precisely why it stayed closed. Fable 5 is the general-release version of that capability level: same class of model, plus the safeguard layer Anthropic says it spent the intervening months hardening.

The timing is not subtle. The release lands days after Anthropic confidentially filed for an IPO, with a reported $47 billion revenue run rate (up from roughly $10 billion a year earlier) and a $965 billion valuation that tops OpenAI's $852 billion from late March. It also lands days after Anthropic publicly called for "a coordinated brake pedal" on frontier AI development — a contrast TechCrunch and NBC's coverage both lean on. Shipping the strongest public model in the world and a warning about models this strong, in the same fortnight, is the whole company in one move.

Is Fable 5 just a scaled-down Mythos?

No — and the distinction changes how you should think about it. "Scaled down" implies a smaller or distilled checkpoint, like Haiku next to Opus. That is not what shipped. Anthropic states Fable 5 and Mythos 5 are the same underlying model; the difference is configuration. Fable 5 runs with safeguards on: classifiers watch for high-risk topics — offensive cybersecurity, biology and chemistry, and attempts to distill the model's outputs — and when one fires, the response is generated by Opus 4.8 instead. Mythos 5 is the same model with those safeguards lifted in specific areas, available only to vetted Project Glasswing partners, with biology researchers named as the next trusted-access group.

In practice: for the 95%+ of sessions that never touch a tripwire, you are running the most capable model Anthropic has ever served, not a lite edition. The honest mental model is Mythos with a circuit breaker, not Mythos junior. Anthropic's confidence number for the breaker: an external bug bounty ran over 1,000 hours of jailbreak attempts against the safeguards and produced no universal jailbreak.

What do the safeguards and 30-day retention mean for developers?

Three things change operationally, and two of them are easy to miss in the launch coverage.

• Silent model substitution — when a safeguard fires, the answer comes from Opus 4.8. If you work in security tooling, exploit analysis, or bio/chem adjacent domains, expect a fraction of responses to be quietly produced by the previous-generation model.
• Mandatory 30-day retention — all Fable/Mythos traffic is retained for 30 days, explicitly including enterprises with existing zero-retention agreements. Anthropic says the data is used only to defend against novel attacks and reduce false positives, never for training. If your compliance posture depends on zero retention, Fable 5 is off the table until that changes.
• A stricter API surface — Fable 5 keeps the Opus 4.8 request shape: adaptive thinking only, and the sampling knobs are gone. temperature, top_p, and top_k all return a 400. One quirk is new: an explicit thinking: { type: 'disabled' } also 400s — omit the parameter entirely if you want thinking off.

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic();

const response = await client.messages.create({
  model: 'claude-fable-5',
  max_tokens: 16000,
  thinking: { type: 'adaptive' }, // the only "on" mode; to disable, omit the field — explicit "disabled" 400s
  messages: [{ role: 'user', content: 'Plan the repo-wide migration in /workspace.' }],
});

// No temperature, top_p, or top_k — all three return a 400 on this model.

The positioning Anthropic pushes hardest is long-horizon autonomy: planning across stages, delegating to sub-agents, and checking its own work over multi-day runs inside an agent harness. If you have been structuring agents along the lines of harness engineering — clear up-front goals, sub-agent fanout, verification loops — this model is built for exactly that shape of work, and the Stripe result is what it looks like when it lands.

How much does Fable 5 cost, and when can you use it?

$10 per million input tokens and $50 per million output tokens. That is exactly 2× Opus 4.8, 3.3× Sonnet 4.6, and less than half what Claude Mythos Preview cost Glasswing participants. The pricing reads as a statement: this is a model for work where the output is worth multiples of the token bill — the Stripe migration would have been a bargain at ten times the price.

Availability: the API has it now as claude-fable-5, and it is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost through June 22, 2026 — from June 23 it draws usage credits. That two-week window is a deliberate try-before-the-meter-starts move; if you want to benchmark it against your own workloads, do it now. Mythos 5 stays restricted to Project Glasswing participants. For wiring any of these models into your own services, the LLM API overview covers the auth and key-handling basics.

When NOT to reach for Fable 5

• Zero-retention compliance — the 30-day retention mandate has no enterprise carve-out. Regulated workloads that require zero retention stay on Opus 4.8 or Sonnet 4.6.
• Security research and offensive-tooling work — this is the domain most likely to trip the safeguards, and a silent Opus 4.8 fallback mid-investigation is worse than knowingly using Opus 4.8 from the start. Glasswing access exists for a reason.
• Everyday IDE and chat traffic — on short tasks the gap to Opus 4.8 shrinks while the bill doubles. Anthropic's own framing concedes this: the lead grows with task length. Sonnet 4.6 remains the sane default driver for interactive coding.
• Latency-sensitive paths — a Mythos-class model at high effort is the wrong tool for autocomplete, classification, or anything a user is actively waiting on. Haiku 4.5 exists for that.
• Determinism-dependent pipelines — with sampling parameters removed there is no temperature: 0 to reach for. If your pipeline assumed it (it never guaranteed identical outputs anyway), re-validate before swapping models.

Frequently Asked Questions

References

• Anthropic — the Fable 5 and Mythos 5 announcement — the announcement: same-model relationship, safeguard design, benchmark claims, pricing, and the Stripe result
• Anthropic — Claude Fable product page — capabilities overview and plan availability
• CNBC — Anthropic releases Mythos-like AI model to the public — SWE-Bench Pro figures, Project Glasswing background, and the IPO/valuation context
• TechCrunch — Anthropic releases Claude Fable days after warning AI is becoming too dangerous — the safety-contrast angle, jailbreak bug-bounty numbers, and the Hex benchmark result
• VentureBeat — Anthropic brings Mythos to the masses — GDPval-AA scores and the enterprise-positioning read
• The Decoder — Major gains in coding and science — benchmark roundup including the Pokémon FireRed run