Anthropic launched Claude Fable 5 today. It's the most powerful model they've made publicly available, and it carries a designation that's worth understanding before you just start using it.
What Mythos Is
Back in April 2026, Anthropic released Claude Mythos Preview. Almost nobody could access it. The rollout was limited to a small group of cybersecurity professionals and critical infrastructure operators through a program called Project Glasswing, and the reasoning was specific: Mythos could find and exploit software vulnerabilities faster than most human security experts. In Anthropic's own testing it wrote a working exploit for a disclosed Windows kernel flaw in 31 minutes.
Mythos sat above Opus in Anthropic's model hierarchy. It was the first tier where Anthropic said the capability level required a fundamentally different access model, not just tighter terms of service. By June 2, the Glasswing rollout had expanded to 150 organisations across 15 countries, still focused on security and critical infrastructure.
Today's the first time the underlying capability has been available to anyone.
What Fable 5 Is
Fable 5 is Mythos with safety classifiers applied. Anthropic chose the name deliberately: "fable" shares its root with "mythos," both tracing back to "that which is told." The underlying models are identical. The safeguards are the whole difference.
In practice, queries that touch cybersecurity exploitation, biology, chemistry, or large-scale model distillation get silently rerouted to Opus 4.8. You still get a response. Anthropic says this happens in fewer than 5% of sessions, so for most use cases you're getting full Fable capability.
The technical architecture behind Mythos hasn't been officially published by Anthropic. Community researchers have proposed designs based on prior research papers, but that's reconstruction, not confirmation.
What It Can Do
The early access feedback is worth looking at, with the caveat that these are controlled conditions from carefully selected partners.
Stripe reported migrating a 50 million line Ruby codebase in a day using Fable 5. Their estimate for the same work by hand: an entire engineering team, two months. Hebbia called it the highest-scoring model they'd tested for senior-level financial reasoning. An investment firm (IMC) said it aced their trading analysis evaluations.
On vision tasks, Fable 5 completed Pokémon FireRed using only raw screenshots, no helper tools. Previous models needed a custom harness for the same thing. Small demo, but it suggests something real about visual reasoning.
Anthropic's own framing: "the longer and more complex the task, the larger Fable 5's lead." This isn't marketed as a chat upgrade. The design bet is sustained, agentic work.
What Changes Practically
Pricing is $10 per million input tokens, $50 per million output tokens. Double Opus 4.8. Through June 22, it's included in Pro, Max, Team, and Enterprise plans. From June 23 onward you'll need usage credits until Anthropic has enough capacity to restore it as a standard feature.
There's also a new data policy. Anthropic now requires 30-day retention on all Mythos-class traffic, including for enterprise customers who previously had zero-retention agreements. The stated purpose is safety monitoring and jailbreak detection, not training. If zero-retention was a hard requirement for your organisation, that has changed.
For workflow and automation work, the meaningful question is whether agentic reliability has actually improved at scale. Multiple early partners called out self-correction and sustained performance over long task chains. I'd test that in your own context before assuming the Stripe benchmark translates directly to what you're building.
What We Don't Know Yet
Real-world performance in general production is unknown. Early access partners were hand-selected and working in favourable conditions.
Pricing and access timeline may shift. Anthropic says it wants Fable 5 back on standard subscription plans as quickly as possible, but there's no committed date.
Mythos 5, the version without cyber safeguards, remains restricted to Project Glasswing partners for now, with a biology research track to come. How that expansion plays out, and how quickly, hasn't been specified.