Anthropic's Claude Fable 5.1 prompting guide: dial down thinking tiers to cut costs, avoid 400 errors

AI寒武纪 · wechat · 2026-09-03

Anthropic published a comprehensive prompt engineering guide for Claude Fable 5.1, addressing behavior changes from the previous generation.

Thinking tiers & cost

Default is high, but Anthropic recommends testing low/medium/xhigh/max on your own eval set: medium roughly matches Fable 5 at notably lower cost; low-tier pricing matches Opus/Sonnet while scoring higher, making it a fit for replacing higher-tier small models.
Low tier calls search tools less; xhigh/max think longer before long outputs.

Engineering practices

Long tool chains produce less user-facing progress by default: enable thinking-block display modes (updates/summarized), remove old rules suppressing output, or explicitly request start/process/final updates.
Batch independent tool calls via a turn-scoped system message asking the model to request all non-dependent items in one response.
History must be append-only: thinking blocks are bound to their exact context on new accounts (created after Aug 31, 2026); mutating the prefix (inserting/deleting reminders, in-place summaries, mid-session system prompt edits) triggers 400 errors. Use turn-scoped messages for dynamic reminders and server-side compaction/context editing for trimming.
When compacting long conversations, explicitly preserve: obstacles and fixes, tried/rejected approaches with reasons, user agreements, progress and open items, and hard-to-reconstruct details (names, numbers, links). Keep user words verbatim, compress the model's reasoning.
Constrain scope: don't fix unrelated bugs (log them instead), resolve ambiguity with the most direct interpretation, keep test files in line with repo conventions.
Prefer high tier; if using xhigh/max, raise maxtokens and forbid drafting full deliverables in the thinking chain to avoid doubled latency or truncation.

Other tips

The model uses less bold/lists, so old anti-formatting rules can be removed or made conditional; add "remove all ornate style" if prose gets too dense.
In autonomous tasks, instruct reversible operations to proceed without pausing for permission.
Combat low-tier search avoidance with a verification reminder for fast-moving domains or uncertain names.
Reduce safety false positives: ask "are there bugs" instead of "does it compile", add docs for obscure languages, strip Base64 from tool outputs.
Prefer local precise edits over full file rewrites; parallelize main and sub-agents; provide crop/zoom image tools for dense charts.

More from coding & agent