Claude Fable 5: Anthropic ships its most powerful model to the public

On June 9 Anthropic released Fable 5, its first publicly available "Mythos-class" model. What it does, what it costs, how to try it, and why it ships with built-in guardrails.

Anthropic shipped Claude Fable 5 on June 9, 2026. It's the most powerful model they've ever made publicly available: the first "Mythos-class" version anyone can use today. For someone like me who builds systems that run on their own, this isn't a routine update. It's the jump that takes agents from working for hours to working for whole days.

Fable 5 and Mythos 5: two names, one brain

Under Fable 5 sits the same architecture as Claude Mythos 5, the frontier model Anthropic keeps restricted to controlled institutions: critical infrastructure across 15 countries, selected biology researchers, partners in the Glasswing program. Fable is the public version. Same brain, with extra guardrails on dual-use capabilities. In practice it's Anthropic's most advanced model, packaged so it can be handed to everyone.

What it can do (and where it really matters)

On paper it was declared state of the art on nearly every benchmark tested. But the numbers that matter if you actually work are these:

• Coding and migrations: Anthropic reports a 50-million-line Ruby codebase migration completed in a single day, against the two-plus months of the classic approach. Stripe talks about "months of engineering compressed into days".
• Long autonomous sessions: inside a harness like Claude Code it can work for days, planning in stages, delegating to sub-agents and checking its own work.
• Vision: it pulls precise numbers from scientific charts, rebuilds a web app's source code from a screenshot, and finished Pokémon FireRed using vision alone, with no helper tools.
• Knowledge work: first model to break 90% on Hex's analytics benchmark, and the highest score ever on Hebbia's Finance Benchmark.

What it costs and how to try it

Pricing is $10 per million input tokens and $50 per million output tokens: roughly double Opus 4.8, and less than half what the Mythos preview cost. The 90% prompt-caching discount stays.

Through June 22 it's included at no extra cost on Pro, Max, Team and seat-based Enterprise plans. From June 23, on those plans, it requires usage credits. On the API and consumption-based Enterprise plans it's fully available right away.

The model holds itself back (and that's deliberate)

The most interesting part isn't the power, it's the brake. Fable 5 runs classifiers that, on three high-risk areas, offensive cybersecurity, dual-use biology and chemistry, and model distillation, block the response and route it automatically to Opus 4.8. Anthropic states that more than 95% of sessions never trigger this fallback, and when it happens the user is told.

The safety numbers are serious: no universal jailbreak found in over 1,000 hours of external bug bounty, and demonstrated resistance across 400-turn evaluations. There's also a 30-day data-retention policy on all Mythos-class traffic. Translated: it's at once the most capable and the most closely watched model Anthropic has shipped.

What changes for those of us building with AI

Aima, the system I run for myself and for clients, runs on Claude. A model that holds up across multi-day sessions without losing the thread, that writes its own tests and verifies its own output, raises the bar on what you can really automate: no longer single tasks, but whole processes left to run on their own. The gap with Opus 4.8 shows up wherever long reasoning is needed.

The flip side is the price and the guardrails. For the vast majority of business workflows you never even touch them, but they're worth factoring in. I'm already testing it on my production agents.

If you want to figure out what you can automate in your company with models like this, the starting point isn't the model but the system: an AI Operating System of agents stitched onto how you actually work. If you want to see how it applies to your case, let's talk.