Features

What Claude Fable 5 can do

Claude Fable 5 (claude-fable-5) is Anthropic's most capable generally available model - the first Mythos-class system released to the public. Four capabilities define the jump from Opus 4.8, and a tightened API surface defines how you build with it. This is the independent guide to all of them, drawn from Anthropic's announcement and the official model documentation.

The four pillars

Agents

Long-horizon autonomy

Multi-day agent sessions that plan across stages, delegate to sub-agents, and check their own work. Stripe migrated a 50-million-line Ruby codebase in a single day; a week-long autonomous run produced a genomics model that beat a Science publication.

Explore long-horizon agents -> Reasoning

Adaptive thinking, always on

Fable 5 is the first Claude where adaptive thinking is the only mode - no thinking config, no token budgets. You steer depth with a single effort parameter, from low to max.

How adaptive thinking works -> Context & Memory

1M context & memory

A million-token context window at standard pricing, 128K output, cheaper prompt-cache minimums, server-side compaction, and a file-based memory tool that tripled performance on long-running tasks versus Opus 4.8.

Context window deep dive -> Safety

Mythos-class safety

Shipped under ASL-3 with three real-time classifiers - cyber, bio/chem, and reasoning-extraction - plus an opt-in fallback to Opus 4.8 that triggers in under 5% of sessions.

Inside the safeguards -> Claude Code

Claude Code agent workflows

How to structure Fable 5 coding sessions around planning, repository inspection, implementation checkpoints, test repair, and final review summaries.

Build better coding agents ->

Supported API features

Fable 5 inherits the streamlined request surface introduced with Opus 4.7 and 4.8, then tightens it further. Everything below applies to claude-fable-5 on the Claude API.

Feature	Status	Notes
Vision (images, PDFs, screenshots)	Supported	High-resolution image input; completed Pokemon FireRed with a vision-only harness
Tool use	Supported	User-defined tools, server-side tools, and SDK tool runners
Memory tool	Supported	File-based, client-side memory; 3x better long-task performance vs Opus 4.8
Compaction	Supported (beta)	Server-side summarization for conversations approaching the 1M window
Task budgets	Supported (beta)	Token countdown the model self-moderates against across an agentic loop
Context editing	Supported (beta)	Prunes stale tool results and thinking from long transcripts
Batches API	Supported	Asynchronous processing at 50% of standard prices
Managed Agents	Supported	Server-managed sessions with Outcomes for rubric-graded long runs
Structured outputs	Supported	`output_config.format` with JSON schema enforcement
Extended thinking (`budget_tokens`)	Not supported	Returns 400 - adaptive thinking is always on; use `effort` instead
Sampling parameters	Not supported	`temperature`, `top_p`, `top_k` all return 400
Assistant prefills	Not supported	Last-assistant-turn prefills return 400 - use structured outputs
Fast mode	Not supported	No fast variant of Fable 5 exists at launch

Migrating from an older model? Read the full migration guide ->

Build with Claude Fable 5

Free in all paid Claude plans until June 22, 2026, and available now on the Claude API, Claude Code, Amazon Bedrock, Vertex AI, and Microsoft Foundry.

Read the API quickstart