Hidden reasoning tokens: what you actually pay for when a model 'thinks'
aieducation
A deep dive into how reasoning tokens are generated, metered, and billed — and why one answer can cost 20x another on the same model.
When a reasoning model answers, most of the work you pay for is invisible. Those hidden notes — the chain of thought the model writes to itself — are metered as output tokens. Understanding how they're generated changes how you budget.
What a reasoning token actually is
Every token the model writes, visible or not, is one forward pass through the network. Reasoning tokens are the scratchpad: the model drafts intermediate steps, catches its own mistakes, and sometimes restarts a line of logic. None of this appears in the final answer, but all of it is billed at the output rate — which is typically 3–5x the input rate.
How effort levels translate to token counts
Effort is a budget cap on hidden tokens, not a quality setting per se. Typical ranges:
- Low: 200–800 reasoning tokens
- Medium: 2,000–4,000
- High: 8,000–15,000
- Max: 30,000+, sometimes hitting the model's context ceiling
The visible answer may be identical across all four — the difference is how carefully the model checked itself before writing it.
Why identical prompts cost different amounts
Reasoning length is dynamic. A trivial question at high effort may use 500 tokens because the model concludes quickly. A genuinely hard question at the same setting may burn the full 15,000. You cannot predict the bill from the prompt alone, only from the prompt-and-task combination.
Making it controllable
Three levers help: set an explicit max-reasoning-tokens cap, downgrade effort for tasks that don't need checking, and log reasoning token counts per request so you can spot runaway costs. The reasoning models vs chat models guide covers when the extra tokens actually pay off.
For help instrumenting cost controls, see the Silverberry AI Platform.