Anthropic published a Claude Fable 5.1 prompting guide with 15 behavior changes

Anthropic published a prompt engineering guide for Claude Fable 5.1 that lists fifteen behavioral differences from Fable 5 and pairs each one with a specific prompt fix, an API setting, or both. It is a problem-solving document: each section starts with a problem a team building on the model would notice, then says what to change.

The fifteen behaviors

Anthropic groups them by where they show up. Effort levels: Fable 5.1 gains capability across all levels, largest at higher settings, and at low it calls search and retrieval tools less often, so runs that relied on that at low effort need an extra instruction. Agent loops: the model may issue one tool call per turn in custom coding agents, bash-and-editor harnesses, and computer use, where Fable 5 issued them in parallel. Progress updates: the model writes fewer user-facing notes during long tool chains, so a user sees no updates from the agent for minutes.

Conversation history: Fable 5.1 requires the message array to stay append-only and returns bound to a different conversation when a harness edits earlier turns. Writing style: prose is longer and denser by default. Formatting: chat replies carry less structure than the content sometimes needs. Retrieval: summaries reproduce source wording without marking it as a quotation. Task completion: turns end before the work is done, or the model asks permission for work the user already requested.

The rest of the list covers compaction summaries dropping constraints, unrequested fixes and extra test files, safeguard false positives that return stop_reason: "refusal" on benign coding requests, whole-file rewrites for small changes, long deliverables at xhigh and max hitting max_tokens, lead agents waiting with nothing to do while subagents run, and vision answers on dense images missing detail.

The prompts and headers Anthropic recommends

Each section has a concrete fix. For tool-call batching, Anthropic gives a one-sentence instruction that goes into a turn-scoped system message: "First privately list what you need next; then request every item that doesn't depend on another's result in this one response." The instruction needs the mid-conversation-system-clear-at-2026-08-21 beta header, so the API clears earlier copies once a later user message exists and only the newest one costs tokens.

For progress updates, Anthropic says to first check whether the client is receiving them at all. The model's short between-call notes come back as progress-update thinking blocks, which are empty under the default thinking.display of "omitted". The fix is to set display: "updates" under the thinking-display-updates-2026-08-18 beta header and render each non-empty thinking block as a status line. Anthropic also warns to audit system prompts for lines like "hold all findings for the final response" that Fable 5 needed and Fable 5.1 does not.

The one that costs teams the most

Whole-file rewrites are the section with the largest dollar impact for a codebase agent, because Fable 5.1's default is to rewrite a file that Fable 5 would have changed with a few edits. Anthropic gives a targeted-edits instruction to add to the system prompt, and notes that the change is worth measuring against a real edit workload before deciding whether to keep it.

The guide reads as a list of things Anthropic learned in the two weeks after Fable 5.1 was released, written down before developers started passing those odd behaviours around as rumours. Every fix is a system prompt line or a header, so adopting them takes minutes and needs no code migration. The test for a team is whether the fix improves their own evals: Anthropic asks teams to rerun tests at each effort level on their own tasks rather than trust the guide's claims.

Source

Prompting Claude Fable 5.1, Anthropic docs. Coverage: Anthropic Shares Power Prompts That Improve AI Results, Roger Montti, Search Engine Journal, 15 September 2026.