What changes on the invoice
DeepSeek is moving its V4 line to a two-tier price. Away from peak, deepseek-v4-flash costs about 0.14 dollars per million input tokens and 0.28 for output, while the larger deepseek-v4-pro sits near 0.435 and 0.87, both carrying a one-million-token context window. Inside two daily windows, 09:00 to 12:00 and 14:00 to 18:00 Beijing time, both input and output rates double. DeepSeek says it will give 24 hours notice before any rate change, and it is retiring the older deepseek-chat and deepseek-reasoner names at 15:59 UTC on 24 July 2026 in favour of the explicit v4 names. Nothing about the model changes hour to hour, only what each token costs.
Why the peak lands on Europe's morning
Beijing runs six hours ahead of Frankfurt in summer, so the afternoon peak of 14:00 to 18:00 in China is 08:00 to 12:00 in Germany, France, Spain, Italy and Poland, and 07:00 to 11:00 in the United Kingdom and Portugal. That is precisely when European teams open their laptops, fire the first batch of the day and stack up interactive calls. The morning Beijing peak of 09:00 to 12:00 lands at 03:00 to 06:00 on the continent, deep in the night when almost nothing runs. The result is a lopsided tax, since European daytime usage collides with the expensive Chinese afternoon while Europe's own cheap hours arrive after lunch, once Beijing has gone off-peak past 18:00 local. Read the clock rather than the price sheet and the surcharge is avoidable.
What an operator does before 24 July
Two dates and one habit. First, repoint any code still calling deepseek-chat or deepseek-reasoner to deepseek-v4-flash and deepseek-v4-pro before 24 July, because the old aliases stop resolving and a silent failure in production is a worse surprise than a planned change. Second, move heavy non-interactive work, nightly summarisation, evaluation runs and bulk extraction, into the European afternoon, which is Beijing off-peak, and keep only latency-sensitive calls in the morning. A team spending a few thousand euros a month on tokens can shave a real slice off the bill by shifting when the load hits, with no change to the model or the code path. The lever is the schedule, and it is entirely yours.
Read next: Zhipu's Free Coding Tool Undercuts Claude Code | Microsoft Bets 2.5 Billion That Tools Are Not Enough
