Pricing
Within 1 hour is 20% off Sync, within 24 hours 50% off. Only completed requests are charged; one not started by the deadline expires unbilled.
Async is in beta
Prices and limits can still change before general availability.
| Window | Price | Applies to |
|---|---|---|
Within 1 hour (1h) | 20% off Sync, 0.8x the Sync rate | Input, cached input and output tokens |
Within 24 hours (24h) | 50% off Sync, 0.5x the Sync rate | Input, cached input and output tokens |
The discount is on every token dimension, cached input included. The Sync rate is the model's own, listed on Models and in Pricing. A model that is not offered for Async has no Async rate.
What is charged
- A completed request is charged, including one that started before the
deadline and finished after it (its result says
late: true). - Anything else costs nothing. A request that never started before the
deadline ends
expired, a request cancelled before it started endscancelled, and a request that failed is not charged. A line the create rejected is never billed. - Billing settles as each group of requests finishes, so a running job's
cost_nano_usdclimbs during the run. It is the sum of what was billed.
Example
A job whose requests use 2,000,000 input tokens and 500,000 output tokens, on a model whose Sync rate is $0.20 per 1M input tokens and $0.60 per 1M output tokens (rates chosen to keep the arithmetic plain):
| Input | Output | Job | |
|---|---|---|---|
| Sync | $0.40 | $0.30 | $0.70 |
| Within 1 hour, 20% off Sync | $0.32 | $0.24 | $0.56 |
| Within 24 hours, 50% off Sync | $0.20 | $0.15 | $0.35 |
The same job in nanodollars, the unit of cost_nano_usd: 560,000,000 on the
1-hour window and 350,000,000 on the 24-hour window.
Budgets and balance
Async spend counts against the key's budget and the
organization's balance like any inference. A create is refused with 402
(insufficient_balance or key_budget_exceeded) as it is for Sync.
Limits and errors
The size limits of an async job, the errors a create can answer with, and the codes a failed request carries in its result.
Dedicated endpoints
Serve a model on capacity reserved for your organization, call it with its private model id, and monitor its requests, errors, latency, and tokens.