Loading lesson…
Extended thinking makes Opus smarter but burns hidden tokens. Here is how to budget it without blowing your bill.
When you enable extended thinking on Opus 4.7, the model generates reasoning tokens before producing its visible answer. Those reasoning tokens are billed at the output rate. A 16k thinking budget on a single call can cost $0.40 all by itself.
| Budget | Extra cost (approx) | Typical win |
|---|---|---|
| 1k | $0.025 | Small quality bump on edge cases |
| 4k | $0.10 | Catches most logical errors |
| 16k | $0.40 | Handles genuinely hard problems |
| 32k | $0.80 | Marginal — diminishing returns |
client.messages.create( model="claude-opus-4-7", max_tokens=2048, thinking={"type": "enabled", "budget_tokens": 4096}, messages=[{"role": "user", "content": prompt}], )Start at 4k. Raise only when you see reasoning truncate mid-thought.Run the same 20 prompts with thinking off and on. If accuracy moves less than 3 points, the budget is wasted. If it jumps 10+ points, you found the right use case.
8 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-modelx-opus-extended-thinking-cost-builders
What is the main idea of "Claude Opus 4.7 — extended thinking cost math"?
Which concept is most central to "Claude Opus 4.7 — extended thinking cost math"?
Which use of AI fits this topic best?
What should a careful learner remember about "Extended thinking is not always better"?
You want to use AI after this lesson. What is the safest next step?
How should AI output about extended thinking be treated?
Name one way to verify an AI answer about extended thinking.
Which action would help you apply "Claude Opus 4.7 — extended thinking cost math" responsibly?