Quick answer
Usage-based pricing charges per token, request, or credit instead of a flat fee. It wins for occasional or spiky work and loses when agents run all day, because a looping run has no natural ceiling. Caps, alerts, and a stop condition turn it into a budget rather than a gamble.
Three billing shapes
| Shape | Predictable | Best for |
|---|---|---|
| Subscription | Monthly cost | Daily use |
| Usage-based | No, varies | Occasional or spiky work |
| Hybrid | Partial | Mixed teams |
Where the risk sits
- Agent runs can loop and spend without producing output.
- Context grows through a session, raising cost per turn.
- Retries are billed even when they fail.
- A missing cap turns a bug into an invoice.
The fix is not to avoid usage-based pricing but to bound it: a spend cap, a per-session stop condition, and an alert at the level where you would intervene.
Choose per workload
- Measure the tokens a normal day consumes.
- Multiply by the rate for your models.
- Compare with the subscription price for the same period.
- If usage is spiky, keep usage-based and set a cap.
- If usage is steady and high, a subscription is usually cheaper.
UsageAtlas records the daily volume that estimate needs. See the API vs subscription guide.
Set a floor and a ceiling
Usage-based pricing is easier to live with when you plan a band rather than a number. The floor is your quietest month, the ceiling is a month with a heavy release, and the difference is the risk you accept in exchange for not paying for idle capacity. Budget the ceiling and report against the floor.
Then check the band against the subscription price for the same period. If the ceiling is consistently above it, the fixed plan is cheaper and the variability was never real. The credit guide covers how the meter converts work into cost.
Questions, answered
Is usage-based pricing cheaper?
For occasional or spiky work, usually. For daily agent use, a subscription with a fixed ceiling is often cheaper and more predictable.
How do I cap a variable bill?
Enable a spend limit where offered, set an alert before the limit, and give agents a stop condition so loops end.
Are retries billed?
Yes. Failed attempts still consume tokens, which is why loops are the most expensive failure mode.