What Is Off-Peak AI API Pricing, and Who Can Use It

Off-peak AI API pricing is a discount applied to tokens processed during the provider's quiet hours, with full rates during busy ones. DeepSeek's published rates are the clearest current example: 15 cents per million uncached input tokens and 60 cents per million output tokens off-peak, with rates doubling during weekday peak windows.

A straight halving of your inference bill is a large number. Whether you can collect it depends entirely on one property of your workload, and most teams get the answer wrong on the first pass.

Why providers do this

Inference runs on GPUs that a provider has already paid for. Those GPUs sit at capacity during the working hours of whichever region dominates traffic, and sit underused the rest of the time.

A provider cannot store unused capacity for later. The marginal cost of serving a request at three in the morning on a half-idle cluster is far below the marginal cost at two in the afternoon on a saturated one. Discounting the quiet hours converts idle silicon into revenue and flattens demand at the same time. Airlines, electricity grids and data centres all reach the same answer for the same reason.

This is also why off-peak pricing tends to be generous rather than token. The provider is not trying to shave a few percent off your bill, it is trying to move your traffic.

The deadline test

Here is the only question that matters: **if this request finished eight hours from now instead of eight seconds from now, would anyone notice?**

Run every workload through it honestly.

The pattern is clear enough once written down. Anything with a human waiting cannot move. Anything feeding a scheduled output or a background process usually can, and teams routinely run that second category during the day purely out of habit, because that is when somebody happened to write the cron entry.

The arithmetic

Take a workload producing 40 million output tokens a month, which is a mid-sized background enrichment job.

At the peak rate of $1.20 per million output tokens, that is $48. At the off-peak rate of $0.60, it is $24. The saving is $24 a month, which is not interesting on its own.

Scale it to 400 million output tokens and the same shift is worth $240 a month, or $2,880 a year, for the work of changing when a scheduled job runs. At that point it is one of the cheapest optimisations available, because it requires no change to prompts, models or output quality.

The honest caveat: for most small builders the absolute saving is small, because the absolute bill is small. The time to care is when a single background workload is a visible line on the invoice. Below that, there are larger wins elsewhere, and reducing API costs covers the ones that usually matter more.

Off-peak is not the only discount, and often not the best one

Three mechanisms reduce the same bill, and they stack differently:

• **Off-peak pricing** discounts by *when*. Requires scheduling control, no change to the request.

• **Batch processing** discounts by *urgency*. You submit a job and collect results later, often at a similar discount. Batch inference covers when that is cheaper.

• **Caching** discounts by *repetition*. If many requests share a long prefix, prompt caching cuts the input side far more sharply than a time-of-day discount does.

Caching usually wins for anything with a large fixed system prompt, and it works during peak hours. Check it first. Off-peak is the right tool when your traffic is genuinely shiftable and your prompts are genuinely varied.

Three things to check before you schedule anything

1. **Whose clock.** Peak windows are defined in the provider's chosen timezone, not yours. Read the documented hours rather than assuming your night is their night.

2. **What counts.** Confirm whether the discount applies to input, output, or both, and whether cached input is discounted separately. The headline number is usually the output rate.

3. **What happens at the boundary.** A long job started just inside the window may finish outside it. Budget for a mixed rate rather than the discounted one.

Then move one job, measure one invoice cycle, and decide from the number rather than the estimate. If you are still sizing the workload at all, estimating tokens for a task is the step before this one.

FAQ

Do all providers offer off-peak pricing?

No, and it is not universal even among those that do. It has been most visible from providers competing hard on price. Check the pricing page rather than assuming.

Does off-peak pricing affect quality?

It should not. You are being charged differently for the same model, not routed to a worse one. If a provider ever does route differently by time, that would be a documented behaviour worth reading about.

Is this the same as reserved capacity?

No. Reserved capacity is a commitment to a volume in exchange for a rate. Off-peak is a time-of-day discount with no commitment.

What is the realistic saving?

Up to half of the affected workload's token cost, applied only to the portion you actually shift. If a quarter of your tokens are shiftable, your total saving is nearer an eighth of the bill than a half.

Sources: the off-peak and peak rates cited come from SiliconANGLE's coverage of DeepSeek's V4.1-Flash pricing.