Anthropic ships Claude Fable 5 and Mythos 5!

Anthropic just announced a new model today — Claude Fable 5 — along with a more locked-down sibling called Claude Mythos 5. Both are “Mythos-class” models, and the short version is: Fable 5 is the one you and I can actually use, while Mythos 5 is a restricted-access variant with some of the safeguards lifted for vetted partners.

So what’s new?

The headline claim is that Fable 5 is state-of-the-art on “nearly all” of the benchmarks Anthropic tested it on. Worth keeping in mind these are their numbers from their launch post, so take them with the usual pinch of salt until the independent evals roll in. With that said, the figures they’re throwing around:

• Coding: Stripe apparently used it to migrate a 50-million-line Ruby codebase in a single day — something they reckon would normally take a team about two months. Anthropic also says it tops Cognition’s FrontierCode eval at medium effort and is 3x more token-efficient on coding tasks than previous Claude models.
• Knowledge work: Top score on Hebbia’s finance benchmark, and trading firm IMC says it “aced their trading-analysis evaluations nearly across the board.”
• Vision: State-of-the-art on vision tasks — pulling precise numbers out of scientific figures, rebuilding web app code from a screenshot, and even finishing Pokémon FireRed with vision-only input (older models needed a bunch of helper tooling for that).
• Long context & memory: Holds focus across millions of tokens. In a fun example, when they gave it file-based memory to play Slay the Spire it performed 3x better than Opus 4.8.
• Benchmarks: A 10-point jump over Opus on complex analytics, and the “first to break 90%” on whatever that eval was.

The memory + long-context stuff is the part that actually interests me. I messed around with local models recently and the single biggest wall I hit was the tiny context window — the moment a task got big, things fell apart. So a frontier model leaning hard into long-horizon, persistent-memory work is the direction I’d want to see.

Pricing

• $10 per million input tokens
• $50 per million output tokens

Anthropic says that’s “less than half the price” of the earlier Mythos Preview. Still firmly frontier-tier pricing, mind you — this is not a “run it all day for fun” model. The token-efficiency gains might soften the blow on real workloads, but output at $50/M adds up fast.

Mythos 5

Mythos 5 is the same class of model but with some safeguards lifted, and it’s not generally available — it’s restricted to vetted partners of Anthropic’s “Project Glasswing”: A trusted-access program for biology researchers and a cybersecurity program.

The reason for the split is capability. On the life-sciences side Anthropic claims Mythos accelerated internal drug-design work by roughly 10x, produced strong candidates for 9 of 14 protein targets, had its molecular-biology hypotheses preferred by scientists ~80% of the time over Opus-class models, and ran autonomous genomics work spanning 138 animal species. That’s exactly the dual-use stuff (gene therapy, bio/chem) you’d want gated behind real access controls.

Availability

Fable 5 is live globally today via the Claude API and consumption-based Enterprise plans, with a staged rollout to subscription plans. Mythos 5 stays restricted to the partner programs for now.

I’ll try to get my hands on it and put it to work on something real — more on that later.

Source

• Introducing Claude Fable 5 and Claude Mythos 5 — Anthropic