Claude Fable 5 is here and It's Mythos for the rest of us

Anthropic didn't pick these names randomly. Fable comes from the Latin fabula — "that which is told." Mythos comes from the Greek for story, legend, and myth. They're the same root word, split across two languages. The model underneath is identical. What separates them is not capability — it's what Anthropic has chosen to let each version say out loud.

Mythos Preview launched in April 2026 through Project Glasswing, a closed program involving cyber defenders, critical infrastructure providers, and a handful of cloud giants including AWS, Microsoft, Apple, and CrowdStrike. It was kept restricted for one reason: it was the most capable model ever built for finding and exploiting software vulnerabilities. Anthropic concluded it was too dangerous to release without controls.

Fable 5 is the answer to that problem. Same model. Classifiers added. Topics flagged. When Fable 5 encounters a query about offensive cybersecurity, biology, chemistry, or model distillation, it routes the request to Claude Opus 4.8 instead. The user gets a response — just not from the full-power system. Anthropic says this fallback hits in under 5% of sessions on average. For everyone else, the experience is effectively indistinguishable from Mythos itself.

This is the key insight most coverage has missed. Fable 5 is not a "weaker" version of Mythos 5. It doesn't have fewer parameters, a reduced context window, or dialed-down reasoning. The only architectural difference is a classifier layer that intercepts certain queries before the main model responds. On the benchmarks where that classifier never fires — which is most of them — Fable 5 and Mythos 5 post essentially identical scores. The few asterisked benchmarks in Anthropic's comparison table (like ExploitBench) show Fable trailing because those are exactly the domains where the classifiers engage.

Anthropic led with software engineering in its announcement, and there's a reason for that. The gap over every competing model is the largest there, and the real-world testimony from Stripe makes the abstract numbers concrete: in a 50-million-line Ruby codebase, Fable 5 completed a codebase-wide migration in a single day. The same task would have taken a full team of engineers more than two months by hand.

That is not an incremental improvement. That is a category shift.

Beyond coding, the knowledge work numbers tell a consistent story. Hex reported Fable 5 as the first model to break 90% on their core analytics benchmark for complex long-running tasks — a 10-point jump over Opus 4.8. Trading firm IMC said it aced their trading-analysis evaluations nearly across the board, including factual lookup, root-cause analysis, and expected-value analysis.

Vision is where the capability leap is easiest to picture. Fable 5 cleared Pokémon FireRed from start to finish using only raw game screenshots — no maps, no navigation aids, no helper harness. Previous Claude models needed an elaborate scaffolding system to attempt that same task. Fable did it with vision alone.

Fable 5 stays coherent across millions of tokens in long-running tasks. Anthropic tested this by having the model play the deck-building game Slay the Spire with access to persistent file-based memory. For Fable 5, that memory access improved performance three times more than it did for Opus 4.8. Fable also reached the game's final act three times more frequently.

This matters for the builder reading this: long-horizon agentic tasks — multi-step pipelines, autonomous code agents, extended research workflows — are exactly where the generational gap between Opus and Mythos-class models shows up. Cursor, GitHub, Cognition, and Vercel all confirmed in early testing that the long-horizon improvement is the most meaningful signal in Fable's release.

The Fable safeguard system is worth understanding in detail, because it will affect some users and confuse others. Here's how it actually works.

Anthropic built a set of classifiers — separate AI systems that run on every request before it reaches Fable 5. If a classifier flags the request as touching cybersecurity offense, biology, chemistry, or model distillation, the response is automatically handled by Claude Opus 4.8 instead. The user is informed when this happens. They still get a response, and Opus 4.8 is still a highly capable model — this is not a refusal.

Anthropic ran an external bug bounty — over 1,000 hours of red-team testing — and found no universal jailbreaks on Fable 5's cybersecurity classifiers. UK AISI made partial progress in a brief initial window, but nothing production-viable. One external partner found Fable 5's cyber safeguards were the most robust of any model tested, including Opus 4.8 and Opus 4.7: it complied with zero harmful single-turn requests across 30 different public jailbreak techniques.

The trade-off is real: the classifiers are conservative, and benign queries in adjacent territory will sometimes hit a fallback. A legitimate security researcher asking about vulnerability disclosure patterns might get routed to Opus 4.8. Anthropic acknowledges this directly and says reducing false positives is the immediate post-launch priority.

The cybersecurity story is the one that got the headlines in April, but the biology capabilities are arguably more consequential for the long run. Anthropic's internal protein design experts used Mythos 5 to accelerate aspects of drug design by around ten times. In one evaluation, Mythos 5 — given protein design and bioinformatics tools but no human assistance — matched or beat skilled human operators at every step a scientist normally handles: choosing binding sites, selecting tools, recovering from failures.

Nine of the 14 protein targets tested yielded strong candidates that are currently under investigation as drug leads. Separately, Mythos 5 conducted autonomous genomics research across millions of cells from 138 animal species, trained a custom ML model on that data, and produced a result that outperformed a recent paper published in Science — while being 100 times smaller.

Mythos 5 is not a separate, more powerful model sitting behind a paywall. It is the same model as Fable 5 with the cybersecurity classifiers removed — meaning vetted Glasswing partners can use the full offensive security capabilities that triggered the original restricted release. The US government is involved in the expansion plan, and Anthropic describes the rollout as collaborative with federal partners.

Glasswing's initial use case has produced early results worth noting: the program has already helped cyber defenders secure critically important software. Anthropic published a report on Glasswing's initial findings separately, and it's worth reading if you work in security infrastructure.

Pricing is $10 per million input tokens and $50 per million output tokens — exactly double Opus 4.8, and less than half what Claude Mythos Preview cost. Prompt caching gives a 90% discount on cached input tokens. Context window is 1 million input tokens with 128K output tokens. Extended thinking is supported.

On the API and consumption-based Enterprise plans, Fable 5 is fully available immediately using the model ID claude-fable-5. For subscription plans, the window is short: free access ends June 22, after which usage credits are required. Anthropic has said it intends to restore Fable 5 to standard subscription plans as quickly as capacity allows — demand is high and infrastructure is under strain.

If you're building with Claude today, the upgrade path is clear. Fable 5 is drop-in compatible — swap the model ID in your API calls and you're using a Mythos-class model. For most use cases — coding agents, knowledge work, document analysis, vision pipelines — the performance gain is unambiguous and the classifier will never fire.

Two things are worth planning for. First, if your product touches cybersecurity, biology, or chemistry in any meaningful way, test your specific queries against the classifier before shipping. The false positive rate is conservative by design, and you want to know where your use case lands before your users find out. Second, plan for the June 22 subscription change: if you're running Fable 5 in a product on a subscription plan, you'll need usage credits after that date.

The longer-horizon question is what the Mythos-class tier means for AI development broadly. Anthropic published a warning alongside this launch — a plea to major AI labs to establish a coordinated brake mechanism on frontier development. They've publicly flagged recursive self-improvement (RSI) as a near-term risk. They're simultaneously releasing what they describe as their most dangerous model to date and asking the industry to slow down. That tension isn't hypocrisy. It's a bet that transparency and controlled deployment beats the alternative of someone else releasing an unchecked version first.

Whether that calculation is right is a genuine open question. For builders, the immediate reality is simpler: the most capable publicly available AI model in the world just became free to use for the next two weeks. Use that time well.