Extended thinking looks like just another dial. Set a budget, let the model cogitate, get a better answer. Easy.
The economics are more interesting than that.
Thinking is a cheap form of compute
Thinking tokens cost less than output tokens on most price sheets, and you can often get the same final-answer quality with 3× the thinking budget and 1× the output budget as you would with 1× thinking and 3× output. That's a real lever if your task tolerates a few extra seconds.
Where it breaks even
Rough rule of thumb from three teams I've talked to:
| Task | Thinking budget | Win over no-thinking |
|---|---|---|
| Code review / refactor suggestions | 8k | ~18% fewer miss-fixes |
| Multi-hop Q&A | 12k | ~25% accuracy lift |
| Structured extraction | 2k | marginal |
| Chat completions | 0–1k | negligible |
The pattern: thinking pays for itself when the task has branches. Extraction doesn't branch. Reasoning over ambiguous evidence does.
Practical advice
- Default to a medium thinking budget for anything involving code or multi-step reasoning.
- Never set a budget without measuring — it's easy to pay for tokens that don't improve output.
- Log the ratio of thinking:output tokens in your traces. It's a leading indicator when something regresses.