What changed
DeepSeek emailed API subscribers on 2026-06-30 that its V4 API will switch to peak and off-peak pricing when the model leaves preview for general availability in mid-July 2026 1. During two daily windows, 09:00 to 12:00 and 14:00 to 18:00 Beijing time, calls bill at twice the off-peak rate 1; for deepseek-v4-pro, output goes from 6 yuan to 12 yuan (about US$1.77) per million tokens 2. This is a reversal: the same lab triggered a price war in May 2026 with a permanent 75% cut on V4 access that forced ByteDance and Tencent to follow 2, and now it adds a surcharge it frames as “better distribution of resources and to enhance service stability” 2.
The clock inside the bill
| Beijing window | Multiplier | deepseek-v4-pro output |
|---|---|---|
| 09:00-12:00 and 14:00-18:00 | 2x 1 | 12 yuan / Mtok 2 |
| all other hours | 1x | 6 yuan / Mtok 2 |
The facts are small; the consequence is not. A daily clock now sits inside the cost of every DeepSeek call, so any router or agent loop that fans out many requests has to reason about the provider’s local time to keep its bill flat. Time-of-day joins model quality and per-token cost as a routing dimension, which is exactly the call an independent gateway analyst already made: routing tables now need a clock 3. A gateway that arbitrates only on a static per-token price will systematically overpay, because the cheapest provider at 03:00 Beijing is not the cheapest at 10:00.
Read the surcharge as a tell, not a footnote. The lab that set the price-war floor is the first to walk it back, which says the discounted floor for serving 1M-context MoE inference 4 was not holding.
Impact on your team
If you route production traffic through deepseek-v4-pro, you have a mid-July deadline, not a decision to defer. Three concrete moves before GA. Shift deferrable batch work such as evals, backfills, and offline enrichment to off-peak Beijing hours, so it never touches the 2x window. Add a peak-window fallback: send latency-tolerant calls to deepseek-v4-flash or a rival during the two windows, and keep deepseek-v4-pro for work that justifies the surge. And make your router time-zone-aware now 3: if it reasons only about model and static price, it will quietly overpay from the day billing starts. What to ignore for now: the speculation that rivals will match the surge 3. None has, so plan around DeepSeek’s clock alone, not an industry-wide one.