🛠 Claude Fable 5.1 and the 'Symptom → Fix' Guide
On September 1, Anthropic released claude-fable-5-1: 1M token context, pricing of $10/$50 per MTok, and cache reads at $0.25. The 'Prompting Claude Fable 5.1' guide notes: the model more often writes 'ornate' text with metaphors, less often comments on agentic chains, and calls tools one by one instead of in batches.
🌍 Minor model versions change the behavior on which agentic systems rely. Anthropic's fixes move from retraining to prompts and API beta headers: thinking.display and turn-scoped system messages.
👤 Fixes from the guide: an anti-ornate fragment in the system prompt, thinking.display: "updates" for statuses, batching independent tool calls, and explicit task constraints. At medium, the quality is roughly the same as Fable 5, but cheaper.
Source 1: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1 Source 2: https://platform.claude.com/docs/en/release-notes/overview