DeepSeek Peak and Off-Peak Pricing Explained: UTC Windows

DeepSeek's API price depends on when you send the request. Two weekday windows in UTC are billed at peak rates, everything else is off-peak, and the gap is wide enough to change which provider is cheapest for you. This guide explains the windows, who lands in them, and how to put a number on it.

All rates and windows below come from our pricing store, which mirrors DeepSeek's models and pricing page. It was last verified on October 13, 2026.

The peak and off-peak windows

DeepSeek bills peak rates Monday to Friday during 01:00–04:00 and 06:00–10:00 UTC, 7 hours per weekday. Two exceptions apply:

• Weekends are entirely off-peak. Saturday and Sunday never hit peak rates.
• Chinese public holidays are off-peak. DeepSeek excludes them from the weekday schedule. Our calculator does not resolve the holiday calendar, so holiday weeks cost slightly less than it estimates.

For orientation, 01:00–04:00 UTC is 09:00–12:00 in China (UTC+8). The 06:00–10:00 UTC window is 14:00–18:00 in China, which is also the European morning: 07:00–11:00 in Central European winter time and 08:00–12:00 in summer time.

What peak hours cost

The schedule applies the same 2× multiple to every line of the bill. Per 1M tokens:

• DeepSeek V4.1 Flash: cache-miss input $0.15 off-peak / $0.30 peak; output $0.60 / $1.20; cache-hit input $0.003 / $0.006.
• DeepSeek V4 Pro: cache-miss input $0.66 / $1.32; output $1.98 / $3.96; cache-hit input $0.022 / $0.044.

DeepSeek's page states it directly: off-peak rates are half the peak rates. What matters for budgeting is the peak share: the fraction of your tokens billed during a peak window.

Which traffic lands in peak

Peak share depends on when your users are active, not where your servers are. Assuming traffic is spread evenly across each group's working hours, the windows translate like this:

• Traffic spread evenly across the whole week: 20.8% peak. That is 7 peak hours × 5 weekdays out of 168 hours, and it is the default in our calculator.
• European weekday business hours (08:00–18:00 local): about 30% in winter time and 40% in summer time, because three or four of the ten working hours fall inside 06:00–10:00 UTC.
• East and Southeast Asian business hours (09:00–18:00 at UTC+8): about 78%. Seven of the nine working hours fall inside the two windows.
• North American business hours: 0% on the East Coast. On the West Coast in winter, the last working hour of Monday to Thursday falls in the 01:00 UTC window. Friday's last hour lands on Saturday in UTC, which is off-peak, so the share stays under 10%.
• Scheduled batch jobs: 0% if you run them outside the windows. Moving a nightly job from 07:00 UTC to 11:00 UTC moves it from the peak rate to the off-peak rate.

These are planning shares. Real traffic clusters around specific hours, so measure the share from your request logs when you can.

How much the peak share changes a monthly bill

Take a mid-sized workload on DeepSeek V4.1 Flash: 500M cache-miss input tokens and 100M output tokens a month.

• Entirely off-peak: $135.00
• Evenly spread (20.8% peak): $163.13 (+20.8%)
• European weekday product, winter (30% peak): $175.50 (+30%)
• European weekday product, summer (40% peak): $189.00 (+40%)
• Asian business-hours product (78% peak): $240.30 (+78%)
• Entirely at peak: $270.00

The same workload on DeepSeek V4 Pro runs from $528.00 off-peak to $1,056.00 at peak, with a European summer product at $739.20.

Because every rate scales by the same multiple, the bill grows in a straight line with the peak share. To run your own numbers, open the DeepSeek Token Calculator and set its Peak traffic share (%) field. For example, this link opens it at 40% for a European summer workload.

How to measure and lower your peak share

You can measure the peak share from logs you probably already have. Export a week of request timestamps with token counts, convert the timestamps to UTC, and add up the tokens that fall on a weekday inside a peak window. Divide by the week's total and you have your real share. Use whole weeks, because a weekday-only sample overstates the share.

Once you know the number, four levers move it:

• Shift deferrable work. Embeddings refreshes, document backfills, evaluation runs and nightly summaries rarely need to run at a specific time. Schedule them for weekends or between 10:00 and 01:00 UTC on weekdays, and they bill at off-peak rates.
• Queue rather than call. For features where users accept a delay, such as report generation or bulk enrichment, put requests in a queue and drain it outside the windows.
• Route by time of day. If European morning traffic is the expensive slice, a router can send peak-window requests to a different model and keep DeepSeek for off-peak hours. Compare both models on the same workload first.
• Raise the cache-hit rate. Cache-hit input is far cheaper than cache-miss input in both windows. DeepSeek's context caching guide says caching is on by default and matches only the prefix of a request. Keep the start of your prompts stable (system prompt and reference material first, the variable request last) and the input bill falls whatever the clock says.

Interactive traffic is the hardest part to move: a user who asks a question at 09:00 in Berlin expects an answer then. For that slice, accept the peak rate in your estimate rather than assuming it away.

Why DeepSeek is no longer an automatic price floor

A common shortcut says: if cost is the only criterion, use DeepSeek. At current prices, two things undermine it:

1. The peak schedule. For European and Asian daytime traffic, a large share of tokens is billed at 2× the off-peak rate.
2. Cheaper competitors. At the prices verified on October 13, 2026, the newest small models from other providers sit at or below DeepSeek's uncached rates.

Here are the cheapest active models in our store for 1M input and 250K output tokens, at uncached rates. Scheduled models are ranked by their all-peak cost, the conservative case, with the off-peak cost alongside:

• Claude Haiku 5.5 (Anthropic): $0.225
• GPT-6 Luna (OpenAI): $0.225
• Mistral Small 4 (Mistral): $0.30
• GPT-5.6 Luna (OpenAI): $0.50
• GPT-5.4 nano (OpenAI): $0.5125
• DeepSeek V4.1 Flash (DeepSeek): $0.30 entirely off-peak, $0.60 entirely at peak
• Gemini 3.1 Flash-Lite (Google): $0.625

Two caveats keep this from being a simple ranking. Claude Haiku 5.5's listed rate applies to requests with prompts up to 100K tokens, so very long single prompts cost more there. And DeepSeek's cache-hit input rates ($0.003 off-peak on V4.1 Flash) are a small fraction of its cache-miss rates. A workload that resends the same long prefix all day, such as a fixed system prompt plus a reference document, can still be cheapest on DeepSeek once real hit rates are measured. Our calculator deliberately uses cache-miss input, so it shows the conservative end.

The practical rule: estimate DeepSeek with your real peak share and your measured cache-hit rate, then compare it on the same workload with the cheapest models from other providers. Don't treat it as the floor without running the numbers.

Rates come from the AICostLabs pricing store, verified against DeepSeek's pricing page. Estimates only; verify provider pricing before you commit a budget.