Fable Brain

Adaptive Thinking and Thinking Mode

6 min read

Adaptive Thinking and Thinking Mode

Adaptive thinking is the always-on reasoning system of Fable 5, tuned only through the effort parameter, and the phrase “thinking mode” names two additional, unrelated harness controls; this note is the vault’s single source of truth for all three layers, and other notes defer to it.

What it is

  • Anthropic’s docs call adaptive thinking “the only thinking mode on Claude Fable 5 and Claude Mythos 5” (https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking, retrieved 2026-07-07). It is always on: the model decides per request whether and how much to think.
  • Three distinct layers share the “thinking mode” vocabulary and must never be conflated:
    1. API layer: adaptive thinking plus the effort parameter in output_config. This is the only layer that actually controls reasoning depth on Fable 5.
    2. claude.ai harness layer: the system prompt scaffold carries a literal <thinking_mode>auto</thinking_mode> tag, System Prompt Export 2026-07, L3818. That tag is the harness’s own directive to the model, not an API parameter.
    3. Claude Code harness layer: /config exposes a “thinking mode” toggle saved as alwaysThinkingEnabled in settings. On Fable 5 this toggle is inert because thinking cannot be turned off (https://code.claude.com/docs/en/model-config, retrieved 2026-07-07).

"thinking_mode" is NOT an API parameter. The corpus tag at L3818 is a claude.ai prompt-scaffold directive, and Claude Code's "thinking mode" is a client-side toggle. Only effort inside output_config changes Fable 5 reasoning behavior at the API.

  • The former > [!gap] callouts in Extended Thinking Budgets on adaptive thinking, effort, and thinking display are closed as of 2026-07-07; that note keeps the context-editing and caching mechanics and defers thinking-mode semantics here.

How it works

  • Always on: adaptive thinking applies whenever the thinking parameter is unset; thinking: {type: disabled} is not supported on Fable 5 or Mythos 5 (adaptive-thinking doc, retrieved 2026-07-07).
  • Effort controls depth: values are low, medium, high, xhigh, and max, default high, set inside output_config. xhigh is available on Fable 5, Mythos 5, Opus 4.8, Opus 4.7, and Sonnet 5 (https://platform.claude.com/docs/en/build-with-claude/effort, retrieved 2026-07-07).
  • Manual budgets are gone: passing a manual budget_tokens value is rejected with a 400 error on Fable 5. Sampling knobs are locked too: non-default temperature, top_p, or top_k are rejected with a 400 error (adaptive-thinking doc, retrieved 2026-07-07).
  • Visibility: raw chain of thought is never returned. thinking.display: "summarized" returns a readable summary; "omitted", the default on these models, returns an empty thinking field (adaptive-thinking doc, retrieved 2026-07-07).
  • Interleaving: adaptive thinking automatically enables interleaved thinking, so the model can think between tool calls without a beta header (adaptive-thinking doc, retrieved 2026-07-07).
  • Extraction refusals: attempts to extract raw reasoning can be refused with stop_details.category set to "reasoning_extraction" (adaptive-thinking doc, retrieved 2026-07-07).
  • claude.ai harness: the export’s scaffold ends with the <thinking_mode>auto</thinking_mode> tag immediately before the Human turn, System Prompt Export 2026-07, L3818, showing the consumer harness delegates thinking depth to the model rather than pinning a budget.
  • Claude Code: the model-config doc states “Thinking cannot be turned off on Fable 5.” and notes that MAX_THINKING_TOKENS=0 turns thinking off on the API for other models but not for Fable 5 (retrieved 2026-07-07).

Best practice

  • Treat this note as the single source of truth for thinking behavior; when another vault note discusses thinking modes, it should link here instead of restating parameters. PRACTITIONER
  • Control reasoning depth on Fable 5 only through effort in output_config; leave the thinking parameter unset. EVIDENCE-BASED
  • Remove budget_tokens, custom temperature, top_p, and top_k from request templates before migrating to Fable 5; each triggers a 400 error. EVIDENCE-BASED
  • Set thinking.display: "summarized" when an application needs to show users a reasoning trace; expect an empty thinking field under the default "omitted". EVIDENCE-BASED
  • Use xhigh or max effort for deep agentic work and low for latency-sensitive routing, keeping the default high otherwise. PRACTITIONER
  • In Claude Code on Fable 5, ignore the alwaysThinkingEnabled toggle when debugging behavior differences; it cannot disable thinking on this model. EVIDENCE-BASED
  • Do not design prompts that ask Fable 5 to reveal raw chain of thought; the API refuses with a reasoning_extraction stop category. EVIDENCE-BASED

Pitfalls

  • Conflating the three layers: sending thinking_mode as an API parameter because the corpus scaffold contains that tag; the API does not accept it.
  • Expecting MAX_THINKING_TOKENS=0 or the Claude Code toggle to silence thinking on Fable 5; both are inert on this model.
  • Carrying budget_tokens forward from Sonnet-era extended thinking code and misreading the resulting 400 error as an outage.
  • Tuning temperature for creativity on Fable 5; any non-default sampling value is rejected, so variation must come from prompting.
  • Logging the empty thinking field under "omitted" display and concluding the model did not think; adaptive thinking still ran.
  • Assuming the summarized display is the raw reasoning; it is a readable summary, and the raw chain of thought is never returned.
  • Citing the L3818 tag as evidence about the API; it is evidence about the claude.ai harness capture only.

Sources

Next actions

  • Re-verify effort defaults and display semantics against the live docs at the next research-pack refresh.
  • Add a comparison note on effort versus legacy budget_tokens cost behavior if Anthropic publishes pricing guidance.

Graph View

Backlinks