Anthropic's biggest leap yet - wrapped in its strictest guardrails
Fable 5 is, by Anthropic's account, the most capable model it has ever made generally available - top of the charts on nearly every benchmark it tested, and pulling further ahead the longer and more complex the task. The twist is the safety architecture bolted on top, and a naming convention worth understanding.
Fable vs. Mythos: same brain, different guardrails
"Mythos-class" is Anthropic's tier above Opus. Fable 5 and Mythos 5 are the same underlying model; the difference is the safeguards. Fable 5 is the public version with protections on. Mythos 5 has the cyber safeguards lifted and goes only to a small group of cyberdefenders and infrastructure providers through Project Glasswing, initially as an upgrade to Mythos Preview.
What it can actually do
The capability claims are eye-catching:
- In testing, Stripe said Fable 5 compressed months of engineering into days - running a codebase-wide migration on a 50-million-line Ruby codebase in a single day, work that would otherwise have taken a team over two months by hand.
- It's the new state-of-the-art on vision, able to rebuild a web app's source code from screenshots alone, and it beat Pokemon FireRed using only raw game screenshots - no maps or helper tools.
- For memory and long-context work, it stays coherent across millions of tokens and improves its own output using persistent notes.
- With Mythos 5, Anthropic's scientists accelerated parts of protein and drug design roughly 10x, and say the model produced novel molecular-biology hypotheses they preferred about 80% of the time in blind comparisons - one later corroborated by an independent lab.
The safety bet: route, don't refuse
Rather than flatly refusing risky prompts, Fable 5 hands them to Opus 4.8. When its classifiers detect requests touching cybersecurity, biology/chemistry, or model distillation, Opus answers instead and the user is told. Anthropic says more than 95% of sessions involve no fallback at all - so for most users, Fable performs like Mythos. It deliberately tuned the filters conservatively, conceding they'll sometimes catch harmless requests, with a promise to narrow false positives over time.
Why so cautious? Because Mythos-class cyber skills could give real uplift to attackers. Anthropic ran a 1,000+ hour external bug bounty that found no universal jailbreak, though it candidly notes the UK's AI Safety Institute made early progress toward one - and that perfect jailbreak resistance probably isn't achievable for anyone.
A new data-retention rule worth flagging
For Mythos-class models, Anthropic now requires 30-day retention of all traffic, on both first- and third-party surfaces, used only for safety rather than training, with logged human access and deletion after 30 days. It's a genuine trade-off with privacy-sensitive customers, but it's how the company plans to detect novel attacks that play out across many requests.
The cliffhanger
Pricing undercuts Mythos Preview by more than half, and Anthropic staged the consumer rollout to manage demand it expected to be very high. But there's an unavoidable epilogue: three days after launch, a US government directive forced Anthropic to suspend access to both models. For now, the most capable Claude you could briefly use is offline - an abrupt pause on one of the company's most ambitious releases.