What DeepSeek changed
DeepSeek confirmed on 30 June that its next model, V4, will arrive in mid-July with a pricing structure no major AI vendor has used before: the API meter runs at two speeds. During peak hours - set to 09:00-12:00 and 14:00-18:00 Beijing time - the per-token rate doubles. Outside those windows, the off-peak rate is the same price DeepSeek charges today. For the flagship deepseek-v4-pro, that is 6 yuan per million output tokens off-peak and 12 yuan at peak; the smaller v4-flash moves from 2 yuan to 4 yuan. Every tier carries a one million token context window, and DeepSeek says customers get an email 24 hours before any rate change takes effect.
The company is careful to frame this as load management rather than a price rise. Off-peak rates do not move, so a workload that already ran overnight pays exactly what it paid before. What changed is that DeepSeek has attached a surcharge to the hours when everyone wants the model at once. That is not a rounding detail. It is the first time a frontier-scale API has told developers that the clock, not just the token count, sets the bill.
Why this is inference priced like electricity
Power grids have charged this way for decades. When demand peaks and supply is tight, the price goes up to push flexible load into quieter hours. DeepSeek is now doing the same thing with GPU time, and the reason is the same: the binding constraint on serving a large model is not the model, it is the compute and the power behind it during the busy window. Flat per-token pricing assumed capacity was effectively unlimited and only volume mattered. Peak-and-off-peak pricing admits the opposite - that capacity is scarce at specific hours, and the vendor would rather pay you, in a discount, to move.
The consequence for buyers is a break with the story that AI simply gets cheaper every quarter. It now gets cheaper or more expensive depending on when you run it. That turns workload scheduling into a cost lever for the first time. Anything that can wait - overnight document processing, batch classification, offline evaluation, retraining data prep - belongs in the off-peak window. Anything interactive and time-sensitive pays the premium. The teams that treat inference like a power bill, and shift the flexible load, will quietly run the same work for meaningfully less than the teams that never look at the clock.
What a European operator should do about it
The timing math favours Europe. Beijing is six hours ahead of Central European Summer Time, so DeepSeek's peak windows land at roughly 03:00-06:00 and 08:00-12:00 in Berlin, Paris or Madrid, and about an hour earlier in London and Lisbon. The practical read: a European business running heavy inference during its own working afternoon, from lunchtime onward, sits almost entirely inside the cheap off-peak rate. The expensive slot for European users is the mid-morning, 08:00 to noon. Off-peak output on v4-pro works out to roughly 0.75 euro per million tokens, against about 1.50 euro at peak - a gap worth engineering around.
The bottom line is to stop treating a token price as a single number. Map your inference load against these windows, push every deferrable job past midday local time, and reserve the morning peak for traffic that genuinely cannot wait. Watch, too, for the pattern spreading: Together AI and others have already repriced their inference layers this year, and once one major vendor proves that time-of-day pricing sticks, the flat rate looks less like a promise and more like a temporary phase. Build the schedule now, while it is still a discount rather than a default.
Read next: OpenAI Buys Its Speed From One Chipmaker | DeepSeek V4 Is Free to Own. Most Companies Will Still Rent It by the Token.
