Extended Thinking & Budget Tokens — Extended Thinking & Adaptive Reasoning

From Manual Budget Tokens to Adaptive Reasoning

Anthropic pioneered Extended Thinking on Claude 3.7 Sonnet (retired Feb 19, 2026) and legacy Claude 4.5 models (claude-haiku-4-5), enabling the model to emit a verifiable internal chain of thought (type: "thinking") before its final answer. Important October 2026 Migration Rule: Starting with Claude 4.7 and across all Claude 5.x models (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5), manual budget_tokens (type: "enabled") is deprecated and rejected with an HTTP 400 error—replaced by thinking: { type: "adaptive" } and output_config.effort.

Legacy budget_tokens Schema (Haiku 4.5 Maintenance Only)

{
  "model": "claude-haiku-4-5",
  "max_tokens": 16000,
  "thinking": {
    "type": "enabled",
    "budget_tokens": 4096
  },
  "messages": [
    { "role": "user", "content": "Write an optimized red-black tree in Rust with full invariant tests." }
  ]
}

Critical API Rules for Thinking Blocks:

• Claude 5.x Standard: Use thinking: { type: "adaptive" } + output_config: { effort: "low" | "medium" | "high" | "max" } (passing budget_tokens or temperature on 5.x throws HTTP 400).
• Legacy Minimum Budget (4.5 only): Where supported on legacy 4.5 endpoints, budget_tokens required at least 1,024 tokens and max_tokens > budget_tokens.
• Signed Thinking Blocks: The API returns thinking inside a thinking block accompanied by a cryptographic signature. In multi-turn tool loops, thinking blocks must be passed back unmodified.