← All posts

Claude Fable 5: Anthropic Ships a Mythos-Class Model to Everyone

A few days ago we wrote about Claude Mythos — the frontier model so capable at finding software vulnerabilities that Anthropic kept it behind a gated, defenders-first program. On June 9, 2026, Anthropic did something that seemed impossible then: it released a Mythos-class model to everyone. It's called Claude Fable 5, and it's the most powerful model Anthropic has ever made generally available — paired with a new safety architecture that makes shipping it broadly defensible. Here's what it is, what it can do, how the safeguards work, and what it costs.

Fable 5 and Mythos 5: same brain, different guardrails

Anthropic launched two models together. Claude Fable 5 is, in their words, "a Mythos-class model that we've made safe for general use." Claude Mythos 5 is the same underlying model with the safeguards removed in certain domains, available only to vetted organizations. The names even encode the relationship: Fable comes from the Latin fabula ("that which is told"), kin to the Greek mythos — the safeguards are the only thing that separates them.

So Fable 5 is the public, safeguarded sibling of the Mythos model we covered earlier. For the >95% of sessions that never trip a safeguard, Fable 5 performs identically to Mythos 5 — you're getting the full frontier model.

What it can do

Anthropic claims state-of-the-art results on nearly every benchmark tested, and the partner anecdotes are the eye-openers:

  • Software engineering. Stripe reported Fable 5 "compressed months of engineering into days" — migrating a 50-million-line Ruby codebase in a day, work that would have taken a team over two months by hand. On Cognition's FrontierCode eval it scores highest among frontier models even at medium effort.
  • Knowledge work. Top score on Hebbia's senior-level finance benchmark; one analytics firm said it's the first model to break 90% on their core long-running analytics benchmark — a 10-point jump over Opus.
  • Vision (new SOTA). It extracts precise numbers from scientific figures and can rebuild a web app's source from screenshots alone. Telling detail: earlier Claude models needed a custom helper harness to play Pokémon — Fable 5 finished the game with vision alone.
  • Long-horizon agentic work. In an agent harness like Claude Code or Managed Agents it can work for days — planning across stages, delegating to sub-agents, and checking its own work. It stays focused across millions of tokens and improves using its own notes (file-based memory helped it 3× more than it helps Opus 4.8).

How much better than Opus 4.8?

Opus 4.8 is still excellent (and, as you'll see, it's Fable's safety net). But Fable 5 pulls ahead, especially as tasks get longer and harder:

  • Beats Opus 4.8 on Anthropic's everyday spreadsheet suite at every effort level, with fewer turns — finishing runs 25–30% faster.
  • A physics-research org called it the strongest model they've tested on frontier physics while using one-third the reasoning tokens — and in 36 hours it got nearly to where GPT-5.5 landed after four days.

The pattern Anthropic emphasizes: the longer and more complex the task, the larger Fable 5's lead.

The safeguards — the reason it can ship at all

Mythos stayed gated because a model this good at exploitation is dangerous in the wrong hands. Fable 5's breakthrough is the safety layer that makes broad release responsible. It works like this:

  • Separate classifiers — independent AI systems — watch for misuse and jailbreak attempts and stop the main model from answering high-risk requests.
  • When a classifier fires, the request falls back to Claude Opus 4.8, and the user is told. As Anthropic puts it, a fallback to a still-excellent model "is a far better experience than an outright refusal."
  • It's rare: safeguards trigger in under 5% of sessions. For everything else, you get full Fable 5.
High-risk domainBehavior
CybersecurityBlocks exploitation / offensive-cyber help → falls back to Opus 4.8. An external tester called the cyber safeguards the most robust of any model, with zero harmful single-turn compliance.
Biology & chemistryFalls back to Opus 4.8 on most requests (dual-use / bioweapon risk).
DistillationAttempts to extract the model's capabilities to train a competitor are routed to Opus 4.8.

⚠️ This is the responsible-disclosure thesis turned into a product. It's the same tension we keep returning to in Security in the Age of AI and Least Privilege for AI Agents: a capability this strong is dual-use, so you wrap it in controls. Anthropic is candid that no system completely stops universal jailbreaks — a 1,000+-hour bug bounty found none, but the goal is to make any that remain "slow and costly" to find. Treat Fable like the powerful tool it is.

What Mythos 5 unlocks (and why it's still gated)

Mythos 5 — Fable with domain safeguards lifted — goes to vetted partners only, and the science results show why both the promise and the caution are real. Anthropic's protein-design experts said it accelerated parts of drug design ~10× (9 of 14 protein targets yielded strong candidates). Scientists preferred its molecular-biology hypotheses ~80% of the time over Opus-class models, and one was later corroborated in peer-reviewed work. In genomics it ran a week of largely autonomous research and produced an ML model that outperformed one published in Science — while being 100× smaller. The flip side: those same capabilities (e.g., predicting viral assembly) are exactly why the biology/chemistry safeguards exist. Project Glasswing participants get the cyber safeguards removed; a trusted-access program for biology is launching for select biomedical researchers.

Pricing & availability

Detail
Price$10 / million input · $50 / million output (2× Opus 4.8 — but less than half of Mythos Preview's $25/$125). On a safeguard fallback you pay Opus prices for the Opus tokens.
API model IDclaude-fable-5 (AWS: anthropic.claude-fable-5 / global.anthropic.claude-fable-5)
PlatformsClaude API, Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, Microsoft Foundry
SubscriptionsFree on Pro/Max/Team/Enterprise Jun 9–22, then usage-credit-based; API + consumption Enterprise available immediately
Data retentionMandatory 30-day retention for all Mythos-class traffic; not used to train models or for non-safety purposes; access logged, deleted after 30 days

What it means for builders and teams

A few practical takeaways:

  • Long-horizon work is now realistic. "Works for days, plans, delegates, self-checks" is the headline — pair Fable 5 with an agent platform (Claude Code, Managed Agents, or AgentCore) and give it a clear goal up front.
  • Price it deliberately. At 2× Opus, Fable 5 is for the hard, high-value tasks where being right and fast pays for itself — not your high-volume cheap path (that's where models like Grok Build or Haiku live). Route by task.
  • Expect the occasional fallback in security/bio/chem work — design for the "you got Opus 4.8 here" case rather than a hard failure.
  • Mind the 30-day retention requirement if you handle sensitive data.

The bottom line

Claude Fable 5 is a genuine milestone: frontier-of-frontier capability — days-long autonomy, SOTA coding and vision, research that rivals published science — made broadly available because the safety layer finally made that defensible. It's also a marker of where we are: the most useful AI is now powerful enough that shipping it requires guardrails, and the gated Mythos 5 sits right behind it for the few who are cleared. For everyone building on top, the skills that turn a powerful model into a safe, productive system — agent design, security, least privilege, and the cloud fundamentals underneath — are exactly what our certification labs are built to teach.

Sources

← Back to all posts