Adaptive Thinking in Series 5 — Extended Thinking & Adaptive Reasoning

The Evolution: Adaptive Thinking & Effort Control

Across the Claude 5.x generation (Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5), Anthropic standardized on Adaptive Thinking (thinking: { type: "adaptive" }). Rather than forcing reasoning on trivial prompts or busting prompt caches when changing manual token budgets, Claude dynamically decides whether and how deeply to reason per turn, steered via output_config.effort ("low", "medium", "high", "max").

How Adaptive Thinking Works in Series 5:

• Autonomous Depth Calibration: Simple queries skip unnecessary thinking for sub-second responses, while complex architectural or mathematical prompts automatically trigger deep multi-branch reasoning.
• Automatic Interleaved Thinking: During tool use, Claude 5.x automatically reasons between tool calls without legacy beta headers (e.g., interleaved-thinking-2025-05-14 is no longer needed). Always pass returned thinking blocks and their signature back unmodified alongside tool_use blocks.
• Always-On Flagships: Both Claude Fable 5.1 (default effort: high) and Claude Opus 5.5 (default effort: medium) keep adaptive thinking permanently enabled for long-horizon reliability.

{
  "model": "claude-sonnet-5-5",
  "max_tokens": 16000,
  "thinking": { "type": "adaptive" },
  "output_config": { "effort": "high" },
  "messages": [
    { "role": "user", "content": "Refactor our authentication middleware to support WebAuthn with fallback passkeys." }
  ]
}