DeepSeek V4 peak billing is weekdays only: 35 hours a week, not 49

Both DeepSeek V4 Pro and DeepSeek V4 Flash stay recommended. What changes is our description of when you pay double: DeepSeek's 2x peak window runs Monday to Friday only. Every hour of Saturday and Sunday bills off-peak.

This is a correction to our own entries, not a vendor price change.

What happened

We re-verified DeepSeek's API rate card on 9 September 2026. The per-token rates are unchanged. The scope of the peak window was not what our entries said.

The rate card states: "Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak)."

Our directory entries for V4 Pro and V4 Flash described "seven daily peak hours". The seven-hour figure is right. The word "daily" was not. Peak pricing never applies on a Saturday or a Sunday.

That is a 14-hour-per-week error, and it ran in the direction that costs you money: we described the expensive window as bigger than it is.

Why it matters

Peak billing covers 35 hours a week, not 49. Of the 168 hours in a week, 133 of them (79%) bill at the off-peak rate.

The practical rule is simpler than the one we gave you. Off-peak runs unbroken for 63 hours, from Friday 10:00 UTC to Monday 01:00 UTC. A batch evaluation, a bulk re-index or a long agent run started on Friday morning never touches the premium rate at all.

Confirmed rates, per 1M tokens (cache miss):

Cache hits bill separately and far lower: $0.022 per 1M off-peak on V4 Pro and $0.007 on V4 Flash, doubling to $0.044 and $0.014 during peak. If your workload replays a large shared prefix, the peak multiplier barely registers on the input side of your bill.

What changes for you

• Move schedulable work to the weekend. Batch jobs, evaluations and backfills bill entirely off-peak, with no scheduling logic beyond "not a weekday".
• On weekdays, avoid 01:00-04:00 and 06:00-10:00 UTC. Both windows are stated in UTC. If your cron runs in local time, convert it rather than assuming your morning matches theirs.
• Do not switch on input price alone. GPT-5.6 Luna undercuts Flash on input ($0.20 against $0.22 per 1M) but charges $1.20 on output against Flash's $0.66 off-peak, and it carries a conditional verdict in our directory. Output-heavy agent work still favors Flash.
• Self-hosting remains the exit. Both models are MIT licensed, so the peak window is an API pricing artifact rather than a property of the model.

What we corrected

Both entries now carry the weekday scope in their pricing context. The verdicts are unchanged at recommended: this correction makes the models cheaper than we described, not worse. The August 2026 price rise still stands, and it still cost Flash its position as the cheapest credible API on input.

FAQ

Do DeepSeek peak rates apply at weekends? No. Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday. Every other hour, including all of Saturday and Sunday, bills at the off-peak rate.

How many hours a week does DeepSeek peak pricing cover? 35. Seven hours on each of five weekdays, out of 168 hours in the week.

How much more do peak hours cost? Double. Peak rates are exactly 2x off-peak on input, output and cache hits alike.

Did DeepSeek change its prices? No. The per-token rates on the rate card are unchanged. What changed is our description of the peak window, not DeepSeek's pricing.