Calibrating Extended Thinking Budgets — Claude Advanced Reasoning

Calibrating Reasoning Depth in Production: Effort Tiers

While retired Claude 3.7 Sonnet required manual budget_tokens tuning (which wasted tokens on trivial queries and invalidated prompt caches whenever changed), Claude 5.x (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5) standardizes on thinking: { type: "adaptive" } paired with output_config.effort.

Recommended Effort Calibration Tiers (Claude 5.x)

Critical Production Constraints

• Dynamic Early Termination: Claude ceases thinking the moment it resolves the problem; you are only billed (at standard output rates) for actual thinking tokens generated.
• Why budget_tokens Fails on 5.x: Passing legacy budget_tokens to Claude 4.7 or 5.5 models returns an HTTP 400 error.
• Allocating max_tokens Headroom: Because max_tokens (up to 128K on 5.x) caps combined thinking + visible output tokens, hitting max_tokens mid-thought returns stop_reason: "max_tokens" without a final answer. Always allocate generous max_tokens headroom (e.g., 16000–64000)!

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

// Steer adaptive reasoning depth via output_config.effort
const response = await client.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 32000,
  thinking: { type: "adaptive" },
  output_config: { effort: "high" },
  messages: [
    { role: "user", content: "Verify whether this lock-free queue is ABA-safe under C++ memory_order_acquire/release." }
  ]
});

const thinkingBlock = response.content.find(b => b.type === "thinking");
const textBlock = response.content.find(b => b.type === "text");

console.log("Stop Reason:", response.stop_reason); // "end_turn" indicates complete solution
console.log("Output Tokens (includes thinking):", response.usage.output_tokens);
console.log("Final Answer:", textBlock?.text);