Pricing
Pricing is based on $0.042 per million input tokens. The number of inference decisions adjusts that price. Each request costs $0.042 ÷ 1M × tokens × max(1, 0.05 × decisions).
Here, tokens is the billable input count returned in usage.input_tokens; decisions counts completed inference operations as described below. Input, state, schema and policy are encoded separately with the o200k_base tokenizer, then their counts are added. Text input is counted as supplied; structured values use compact JSON with sorted keys. When a request uses schema_id or policy_id, the stored definition is counted.
Requests with up to 20 decisions are charged at the input-token rate. Beyond that, each additional decision adds 5% of the input-token price. The resolve response does not expose the decision count or the calculated price.
Prices are in US dollars (USD).
Example request pricing
| Input tokens | Decisions | Multiplier | Price (USD) |
|---|---|---|---|
| 1,000 | 10 | 1 | $0.000042 |
| 1,000 | 40 | 2 | $0.000084 |
| 1,000,000 | 100 | 5 | $0.21 |
These are illustrative calculations from the formula, not observed bills.
What counts as a decision
| Operation | Decision count |
|---|---|
| Selecting a value, such as an enum or boolean | One per inference selection |
| Generating a numeric value | One per value-generation operation |
| Selecting an exact source quote | One per quote-selection operation, regardless of the quote's length |
| Generating free-form text | One per 20 decoded Unicode characters, summed across generated text and rounded up |
| Fixed values and deterministic structure | No inference decisions |
The count includes completed operations during internal search, not just the final JSON. A quoted string is charged as a quote selection, not as free-form text. Fixed values still contribute to the input schema's token count, and the base input-token rate still applies.