Since August 16, 2026, DeepSeek bills by time of day. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday; the other 17 hours of each weekday, and all of the weekend, are off-peak and cost exactly half. Same model, same API, same token, two prices, depending on when the request arrives.
That makes scheduling a pricing decision. A workload moved from peak to off-peak gets an automatic 50 percent discount without changing a line of code beyond when it runs. This page is the practical reference: the hours in your time zone, who the peak windows actually hit, the savings math, and what to reschedule. For the rates themselves, see DeepSeek API pricing for every model and host, plus the DeepSeek V4.1 Flash and V4 Pro pages. V4.1 Flash launched September 10, 2026 on the same peak and off-peak schedule.
The hours, by time zone
Peak windows are fixed in UTC: 01:00-04:00 and 06:00-10:00, Monday through Friday. Weekends are off-peak all day. Local equivalents (summer time where applicable):
| Time zone | Peak window 1 | Peak window 2 |
|---|---|---|
| UTC | 01:00 – 04:00 | 06:00 – 10:00 |
| US Eastern (EDT) | 9:00 PM – 12:00 AM | 2:00 AM – 6:00 AM |
| US Pacific (PDT) | 6:00 PM – 9:00 PM | 11:00 PM – 3:00 AM |
| Central Europe (CEST) | 3:00 AM – 6:00 AM | 8:00 AM – 12:00 PM |
| India (IST) | 6:30 AM – 9:30 AM | 11:30 AM – 3:30 PM |
| China (CST) | 9:00 AM – 12:00 PM | 2:00 PM – 6:00 PM |
The UTC windows are fixed, so local times shift by an hour when daylight saving changes; US Eastern peak becomes 8:00-11:00 PM and 1:00-5:00 AM in winter.
Look at the China column and the design is obvious: peak hours are exactly Chinese business hours, 9 to noon and 2 to 6 in Beijing. This is congestion pricing for DeepSeek's domestic daytime demand. Everyone else's bill depends on how their working day happens to overlap with China's.
Who the peak windows actually hit
US teams mostly got lucky. The American working day (9 AM to 5 PM Eastern is 13:00 to 21:00 UTC) sits entirely in off-peak. Interactive traffic from US users bills at half rate all day without anyone doing anything.
The US overnight cron job did not get lucky. The classic 2 AM Eastern batch run lands squarely in the second peak window (2:00-6:00 AM EDT). The habit of pushing heavy jobs to "cheap overnight hours" now does the opposite on DeepSeek: your quietest local hours are the most expensive time of day. Moving that job to anywhere between 6 AM and 9 PM Eastern halves its cost.
European mornings are peak. 8 AM to noon CEST bills at full rate; the afternoon is off-peak. A European team that runs its heavy evals after lunch pays half what a morning-run team pays.
Chinese and Indian daytime workloads carry the premium. Business-hours traffic from Asia is what the pricing is built around, and there is no free schedule adjustment available; for those teams, the decision is model choice and caching, not timing.
The savings math
The full rate schedule:
| V4 Pro off-peak | V4 Pro peak | V4.1 Flash off-peak | V4.1 Flash peak | |
|---|---|---|---|---|
| Input (cache miss) $/1M | $0.66 | $1.32 | $0.15 | $0.30 |
| Input (cache hit) $/1M | $0.022 | $0.044 | $0.003 | $0.006 |
| Output $/1M | $1.98 | $3.96 | $0.60 | $1.20 |
For a workload of 10M input and 1M output tokens per day on V4 Pro:
| Schedule | Daily | Annual |
|---|---|---|
| Entirely at peak | $17.16 | $6,263 |
| Entirely off-peak | $8.58 | $3,132 |
Same workload, same tokens, $3,131 a year apart on timing alone. A workload spread evenly across the week hits peak for 35 of 168 hours, about 21 percent, for a blended $10.37 a day. Add a 60 percent cache hit rate and the off-peak figure drops to $4.75 a day. Stacking the two discounts, off-peak plus cache, gets the bill to roughly a quarter of the peak uncached rate, which is the cheapest way to run a DeepSeek flagship today. The V4 Pro column applies until September 14, 2026, when deepseek-v4-pro requests start routing to V4.1 Flash. The same 10M/1M workload on V4.1 Flash runs $4.20 a day entirely at peak and $2.10 entirely off-peak, so the 50 percent schedule discount carries over unchanged.
What to move off-peak
The candidates are anything that doesn't have a human waiting on it:
- Batch inference and backfills. The obvious wins: summarization queues, document processing, classification runs.
- Evals and regression suites. Continuous eval pipelines are often the biggest silent token spend; they rarely care what hour it is.
- Synthetic data generation. Pure batch, fully schedulable, often enormous volume.
- Agent workflows with deadlines rather than latency needs. Anything due "by morning" can run in the 17 cheap weekday hours, and anything due Monday can run all weekend at half price.
What can't move: interactive traffic. If your users are active during your peak overlap, that premium is the cost of their working hours, and the levers left are caching, model tier (V4.1 Flash at $0.15 and $0.60 off-peak), or a different host, since third parties still bill flat rates around the clock.
Why time-of-day pricing exists
Because inference is a utility, and utilities have always priced this way. Electricity costs more at 6 PM than 3 AM for the same reason DeepSeek tokens now cost more during Beijing business hours: the hardware is finite, demand is lumpy, and a price signal moves the movable load to the empty hours. The economics underneath are GPU utilization economics: a fleet that serves its peak has idle capacity the rest of the day, and half-price tokens are how you sell it.
It is also one more step in tokens behaving like a traded commodity. Power markets went from flat tariffs to time-of-use pricing to day-ahead auctions and forward curves. Tokens now have posted prices that vary by hour; the exchange mechanics that complete that sequence, market-set prices and hedgeable forwards, are the part still being built. A price that changes with the clock, set by announcement, is exactly the kind of price a market eventually takes over.
Frequently asked questions
What are DeepSeek's peak hours?
01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. That is seven hours a day on weekdays, 35 hours a week, and weekends are entirely off-peak.
How much cheaper is DeepSeek off-peak?
Exactly half. Every rate, input, output, and cache hit, on both V4.1 Flash and V4 Pro, is 50 percent of the peak rate during the other 17 weekday hours and all weekend.
When did DeepSeek start peak and off-peak billing?
August 16, 2026 at 16:00 UTC, as part of a broader price increase of 50 to over 1,100 percent depending on token type.
Which price do I pay if a request spans the boundary?
Billing follows when the request is processed, so schedule batch jobs with a margin rather than starting them at the boundary minute.
Do third-party DeepSeek hosts have peak pricing?
No. Third parties list flat rates around the clock. After the August change, most repriced upward on V4 Pro but held DeepSeek's old price on V4 Flash. On V4.1 Flash, launched September 10, the three third-party hosts listed at DeepSeek's peak rate or above on day one, so the official API was the cheapest host at every hour. The full host tables are on DeepSeek API pricing.
Why did DeepSeek pick these hours?
The windows map exactly to Chinese business hours (9:00-12:00 and 14:00-18:00 Beijing time). It's congestion pricing for their domestic daytime demand.
Methodology
Peak windows, the Monday through Friday schedule, and rates from DeepSeek's official API documentation, effective August 16, 2026. Time zone conversions use summer offsets where applicable; UTC windows are fixed year-round. Workload examples use stated token volumes; recompute with your own mix. V4.1 Flash rates verified against the same documentation on September 10, 2026. Last verified: 2026-09-10.