Anthropic's Claude Fable 5 doesn't want a conversation. It wants a deadline.
The original AI interaction was simple: you type, it responds. Even as models grew smarter, humans stayed in the loop — correcting, steering, approving. Fable 5, launched June 9, breaks that contract entirely.
The engine behind this is Adaptive Thinking — always-on in Fable 5, it scales reasoning depth to match task complexity and, crucially, validates its own outputs before calling work done. Pair that with a 1-million-token default context window and the model can hold a sprawling codebase in mind for days without losing the thread.
On Cognition's FrontierCode benchmark — testing real production-quality code, not toy problems — Fable 5 leads all frontier models. GitHub, Cursor, and IMC all reported results in early testing that exceeded any prior Claude model. Legal teams found its contract redlines matched or beat their previous best model in blind review.
Against its rivals, Fable 5 holds a clear lead in long-horizon coherence. GPT-5.5 remains competitive on real-time voice and multimodal tasks, and edged Fable 5 on the new Agents' Last Exam benchmark when paired with OpenAI's Codex orchestration layer. The race is genuinely contested — but the terrain has shifted. Nobody is arguing about MMLU scores anymore. The question is: which model can you trust to run a migration while you sleep?
For now, Anthropic has an answer. A more powerful version — Claude Mythos 5, with safety restrictions partially lifted — remains locked behind Project Glasswing, available only to vetted government partners. What the public gets is already enough to rewrite how engineering teams work.
Claude Fable 5 is available on the Anthropic API, AWS Bedrock, Vertex AI, and Azure Foundry at $10/M input · $50/M output tokens.