The Evolution: Adaptive Thinking

In the Claude 5 generation (Claude Sonnet 5, Claude Opus 5, and Claude Fable 5.1), Anthropic introduced Adaptive Thinking by default. While Claude 3.7 Sonnet required developers to manually allocate and tune a rigid budget_tokens parameter, Series 5 models dynamically analyze problem complexity and modulate internal reasoning depth autonomously per turn.

How Adaptive Thinking Works in Series 5:

• Autonomous Depth Calibration: For simple classification or routine code formatting, the model responds with minimal reasoning overhead. For multi-step architectural refactoring or mathematical synthesis, it automatically allocates deep reasoning steps.
• Fluid Tool Interleaving: Adaptive thinking enables Claude 5 models to think, invoke tools, inspect tool outputs, and resume reasoning without losing chain-of-thought coherence.
• Always-On in Claude Fable 5.1: On flagship models like Claude Fable 5.1, adaptive reasoning is perpetually active, maintaining long-horizon goal state across multi-hour agent sessions.

{
  "model": "claude-sonnet-5",
  "max_tokens": 8192,
  "messages": [
    { "role": "user", "content": "Refactor our authentication middleware to support webauthn with fallback passkeys." }
  ]
}