0%
Anthropic · 2026 · Release 05

Claude
Fable 5

A model that reasons in arcs, not tokens. Fable 5 blends long-horizon planning with the cadence of great storytelling — built for agents that must remember why, not only what comes next.
FBL-5 · 2.0T params · 1M ctx
Scroll to begin
— : — : —
01 — Manifesto

Every model so far has been a sentence. Fable 5 is a story — a model that holds beginnings, middles, and ends in a single thought, weaving plans across hours of reasoning without losing the thread of why it began.

Five chapters in one release.

Fable 5 isn't an incremental step. It rethinks how a language model plans, remembers, and collaborates with agentic tools across long horizons.

01 / Narrative Reasoning

Thinks in arcs, not tokens.

Fable 5 organizes its reasoning into chapters — high-level intents, supporting beats, and resolution checks — letting it sustain coherent goals across thousands of steps without drift.

02 / Persistent Memory

Remembers why, not just what.

A new working-memory layer encodes intent, not just facts — surviving across sessions, tools, and agents.

03 / Agentic Loops

Self-revises in scenes.

Built-in plan → act → critique cycles let Fable 5 catch its own mistakes before you do.

04 / Multimodal Author

Writes worlds, not posts.

Native generation across prose, code, image storyboards, and structured documents — coherent across hours of context.

05 / Tool-Native

Orchestrates tools like a conductor.

Fable 5 was post-trained on millions of agentic traces. It selects, sequences, and recovers from tool calls with the discipline of a senior engineer — making it the most reliable model yet for autonomous workflows.

— A note from the team
"We didn't want a faster parrot.
We wanted a model that could hold a thought long enough to finish one."
Fable 5 Research Lead · Anthropic

Pick a prompt. Watch it think.

Each response is hand-crafted to demonstrate Fable 5's voice — narrative coherence, structural awareness, and metaphorical reach.

Prompt library
04 — Benchmarks

State of the art across reasoning, code, and long-context recall.

Narrative Coherence (1M ctx)

0%
Fable 5Previous

Agentic SWE-bench Verified

0%
Fable 5Previous

GPQA Diamond (reasoning)

0%
Fable 5Previous

Tool-use reliability (50-step)

0%
Fable 5Previous

A library of minds, indexed by intent.

Fable 5's mixture-of-experts routes each thought to specialized "narrators" — reasoning, code, multimodal, and tool-use heads — while a planning conductor binds them into a single coherent voice.

  • 2.0T total · 240B active parameters
  • 1,048,576 token context window
  • 32 expert narrators · 4 active per token
  • Trained on 18T tokens of curated text + traces
  • Constitutional AI v4 alignment
Generally available · June 2026

Begin the next chapter.

Request API access