above.devBuy credit
Back to blog

30 August 2026

DeepSeek peak and off-peak pricing, explained

DeepSeek bills different rates by time of day, defined in Beijing time, with weekends entirely off-peak. Here is exactly when the cheap hours are in UTC and how to schedule batch work around them.

DeepSeek's API bills different rates depending on when your request arrives, and since August 2026 the schedule includes a weekend rule. The details are simple once you see them in UTC, but most summaries get the weekend boundary wrong because DeepSeek defines it in Beijing time.

The schedule

  • Peak: 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, on weekdays only. These windows are Chinese business hours (09:00 to 12:00 and 14:00 to 18:00 Beijing time), which is why they look odd in UTC.
  • Off-peak: every other hour, at half the peak rate.
  • Weekends: entirely off-peak. DeepSeek's weekend is the Beijing weekend, which runs from 16:00 UTC on Friday to 16:00 UTC on Sunday. If you assume the UTC weekend you will mis-price 8 hours on each end.

Off-peak covers 128 of the 168 hours in a week. Unless your traffic is concentrated in those two weekday-morning UTC windows, most of what you send bills at the cheaper rate.

Current rates through above.dev

above.dev sells DeepSeek at the upstream rate plus a flat 10%, and the peak and off-peak split passes straight through. Per million tokens:

  • DeepSeek V4 Flash: off-peak $0.242 input, $0.0077 cache hit, $0.726 output. Peak is exactly double.
  • DeepSeek V4 Pro: off-peak $0.726 input, $0.0242 cache hit, $2.178 output. Peak is exactly double.

The switch is automatic and applies per request at the moment it arrives. You do not opt in, and every response tells you what you actually paid via the x-cost-usd header.

Practical scheduling

If you run batch or overnight agent workloads, the arithmetic is worth a cron entry. A job that runs at 02:00 UTC on a Tuesday pays double what the same job pays at 12:00 UTC, or at any time on the weekend. Shifting bulk evaluation runs, index building, or dataset generation to the Beijing-defined weekend halves their token cost with no code changes.

One caveat for long streams: billing is settled per request, so a request that starts just before a boundary is priced at the rate in effect when it began. For most workloads this rounds to nothing; for very long batch streams it is worth knowing.