Now generally available · Mythos-class

The first Mythos-class model Anthropic has made safe for everyone. State-of-the-art reasoning, vision and memory — and the longer the task, the further ahead it pulls.

Scroll
01 — The leap

A new class of model,
not just a new number.

Fable 5 is the first Mythos-class system released to the public — Anthropic's highest capability tier, built for sustained, multi-step work rather than one-shot answers.

N°01
🧠

Long-horizon reasoning

Plans, executes and refines across many stages while holding context. Its lead over earlier Claude models grows as tasks get longer and harder.

multi-step agents
N°02
👁️

Vision, genuinely solved

Reads precise numbers from scientific figures, rebuilds a web app's source from screenshots alone — and beat Pokémon FireRed with a vision-only harness.

SOTA vision
N°03
♾️

Million-token memory

Stays focused across millions of tokens in long-running tasks, taking and re-reading its own notes to keep improving its output.

1M context
N°04
🎚️

Adaptive thinking only

The old knobs — temperature, top_p — are gone. Instead, one effort dial trades depth for speed, from low to max.

effort: low→max
N°05
🛡️

Safe by design

New classifiers route high-risk requests; sensitive cyber/bio/chem queries get refused or handed to Opus 4.8. In testing, 95% of sessions stayed fully on Fable.

classifier routing
N°06

Built for builders

Up to 128K output tokens, a 90% discount on cached prompt content, and native availability across enterprise plans and developer platforms.

128K output
02 — The lead grows

The harder the task,
the bigger the gap.

Fable 5 is state-of-the-art on nearly every tested benchmark. These bars animate to illustrate the widening lead as task complexity climbs — short prompts are close, long-horizon work isn't.

One-shot Q&Alow complexity
0
Knowledge workmedium
0
Software eng.high
0
Long-horizon agentsvery high
0
Fable 5 Previous generation · Illustrative composite — directional, not exact scores.
03 — One dial to rule them

Steer thinking with effort.

Drag the dial. Fable 5 replaces temperature & top_p with a single adaptive-thinking effort parameter — trading depth against latency and cost.

MED
effort
Reasoning depth▮▮▯▯
Latencyfast
Relative cost1.0×
Best forquick answers

Medium effort. A balanced default — enough deliberation for most everyday tasks without paying for deep chains of thought.

04 — It doesn't lose the plot

A million tokens,
still in focus.

Watch the read-head sweep the context window. Fable 5 sustains attention across long documents and tasks — and writes notes to itself to stay coherent.

0token context window
0max output tokens
0cache input discount
0sessions fully on Fable
05 — Race to the top

Capability, with a conscience.

Try a request. Fable 5's classifiers inspect intent: ordinary work flows straight through; sensitive cyber / bio / chem requests are refused or routed to the more conservative Opus 4.8.

Your request
Refactor my web app
Safety classifier
inspects intent
✓ Handled by Fable 5 — full capability
06 — Two faces of one model

Fable & Mythos.

Same underlying brain, different doors. The difference isn't raw capability — it's how each one responds to high-risk requests.

Claude Fable 5

Public · paid plans & API
  • For everyone. Pro, Max, Team, Enterprise & developer platforms.
  • Guardrailed. Classifiers + fallback to Opus 4.8 on sensitive topics.
  • State of the art on nearly all public capability benchmarks.
  • $10 / $50 per million input / output tokens.

Claude Mythos 5

Restricted · Project Glasswing
  • Same core model — safeguards lifted in some areas.
  • Trusted access only: vetted security & research partners.
  • Cyber-grade: the capability that was deemed too risky to ship openly.
  • Distributed carefully through controlled-access programmes.
0token context
0output tokens
4effort levels
0stay on Fable
The end of one-shot.

Write the next chapter.

Fable 5 is available now to Claude Pro, Max, Team and Enterprise — and across the developer platforms you already use.

Start with Fable 5 →