Fable Brain

Extended Thinking Budgets

5 min read

Extended Thinking Budgets

Token-denominated thinking budgets are a pre-Fable-5 control: on Fable 5 adaptive thinking is always on, budget_tokens is rejected with a 400 error, depth is steered with the effort parameter, and thinking blocks still spend from the same context window that editing and caching manage.

What it is

  • Extended thinking is the API-side reasoning spend: tokens the model uses to think before answering, surfaced as thinking blocks in the conversation and managed like any other context spend.
  • On earlier models the spend was set manually via thinking: {type: "enabled", budget_tokens: N}. On Claude Fable 5 and Claude Mythos 5 that mode is gone: adaptive thinking is the only mode, it cannot be disabled, and a manual budget_tokens request is rejected with a 400 error (https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking, retrieved 2026-07-07).
  • Depth is controlled by the effort parameter in output_config, with values low, medium, high, xhigh, and max; the default is high (https://platform.claude.com/docs/en/build-with-claude/effort, retrieved 2026-07-07).
  • Full thinking-mode semantics (display behavior, harness toggles, interleaved thinking, the corpus thinking_mode tag) live in Adaptive Thinking and Thinking Mode, the anchor note this one defers to; this note owns the budget, editing, and caching mechanics.

How it works

  • Adaptive replaces budgets: whenever the thinking parameter is unset on Fable 5, the model decides its own thinking depth, scaled by effort rather than a fixed token cap (https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking).
  • The raw chain of thought is never returned; thinking.display returns a readable summary when set to summarized, and an empty thinking field under the default omitted (same source).
  • Context editing includes a thinking-specific edit type: clear_thinking_20251015 controls how many thinking turns are preserved when context is cleared (https://platform.claude.com/docs/en/build-with-claude/context-editing.md).
  • When clear_thinking_20251015 is combined with clear_tool_uses_20250919, it must be listed first in the edits array.
  • Context editing executes server-side before the prompt reaches Claude, so the client’s stored history keeps the full thinking record even after edits.
  • Caching interplay: clearing invalidates cached prefixes at the clearing point, while kept thinking blocks preserve cache hits, so preserved thinking turns are also preserved cache economics.
  • Request compatibility: a max_tokens 0 cache pre-warm request is rejected when extended thinking is enabled, alongside streaming, structured outputs, and batches.
  • Budget framing: Anthropic treats context as a finite attention budget targeting the smallest set of high-signal tokens, and thinking tokens draw from the same window as everything else.

Best practice

  • Steer Fable 5’s reasoning depth with effort (default high; xhigh and max for the hardest work), not with budget_tokens, which the API rejects. EVIDENCE-BASED
  • Preserve recent thinking turns when configuring context editing, both for reasoning continuity and to keep cache hits. EVIDENCE-BASED
  • List clear_thinking_20251015 before clear_tool_uses_20250919 whenever both edit types are combined. EVIDENCE-BASED
  • Do not attempt max_tokens 0 cache pre-warming on requests with extended thinking enabled; the API rejects the combination. EVIDENCE-BASED
  • Account for thinking tokens inside the attention budget when sizing long agentic runs. PRACTITIONER
  • Re-verify thinking parameter names against the live platform docs before writing integration code; the budget-to-adaptive migration shows they change between generations. PRACTITIONER

Pitfalls

  • Sending budget_tokens or thinking: {type: "disabled"} to Fable 5; both are unsupported and the budget form returns a 400 error.
  • Expecting raw chain of thought in responses; it is never returned, and prying at it can trigger a reasoning_extraction refusal (see Adaptive Thinking and Thinking Mode).
  • Ordering the edits array wrong; the docs require the thinking edit type first when combined.
  • Clearing thinking aggressively to save space and silently losing cache hits at the clearing point.
  • Combining cache pre-warm requests with extended thinking and misreading the rejection as a caching bug.
  • Assuming thinking tokens are free context; they occupy the same finite window as inputs and outputs.

Sources

Next actions

  • Watch the adaptive-thinking and effort docs for changes at the 2026-08-06 refresh (new effort values or display modes).
  • Probe thinking.display summarized output on a live Fable 5 call and file the observed shape.
  • Confirm clear_thinking_20251015 behavior is unchanged on Fable 5 at next refresh; the editing docs predate the Claude 5 family.

Graph View

Backlinks