The email came three weeks before the price
On 29 June 2026 DeepSeek emailed developers and posted a notice on its open platform: the official V4 release would arrive in mid-July, and it would arrive with two prices instead of one. The release landed on 19 July. The pricing page in the developer documentation still lists a single flat rate per model.
That gap between what was announced and what is published is the whole story for anyone running production traffic through the API. DeepSeek has said it will give 24 hours of notice before an adjustment takes effect. Twenty-four hours is enough time to send an email. It is not enough time to move a workload.
Two prices, one clock, and the clock is in Beijing
The announced peak windows are 09:00 to 12:00 and 14:00 to 18:00 Beijing time, and calls inside them cost twice the off-peak rate. In yuan, V4 Pro output goes from 6 to 12 per million tokens and V4 Flash output from 2 to 4. The published flat rates are 0.87 dollars per million output tokens for Pro and 0.28 for Flash, with cache-miss input at 0.435 and 0.14, which puts Pro output at roughly 0.75 euros or 0.65 pounds per million tokens before any surcharge.
Beijing runs six hours ahead of Frankfurt in summer and seven ahead of London. The second peak window is therefore 08:00 to 12:00 in Frankfurt and 07:00 to 11:00 in London, so a European team pays the surcharge across its own working morning and gets the off-peak rate for the entire afternoon. The first window, 09:00 to 12:00 in Beijing, is 03:00 to 06:00 in Frankfurt, which is precisely where most European shops park their overnight batch.
The deadline you can see is the small one
The documentation carries one hard date. The names deepseek-chat and deepseek-reasoner stop resolving on 24 July 2026 at 15:59 UTC. Both were aliases for the non-thinking and thinking modes of what is now deepseek-v4-flash, so the remedy is a string change in a model field and nothing more.
That is the deadline every team will act on, because it is visible and it breaks loudly. The pricing clock breaks quietly. Nothing fails, no alert fires, and the invoice simply doubles for part of the day on a business calendar set in a market the buyer does not operate in.
What an outside evaluation found under the price
The US Center for AI Standards and Innovation published its own evaluation of DeepSeek V4 Pro, and its numbers do not match the company's. CAISI put V4 Pro at an Elo of 800 against 1260 for GPT-5.5 and 749 for GPT-5.4 mini, judged the model roughly eight months behind the frontier, and measured 46 percent on the held-out ARC-AGI-2 set against 79 percent. It also found that the self-reported results had been presented selectively.
The same evaluation found V4 cheaper than the comparable American model on five of seven benchmarks, in a band running from 53 percent less expensive to 41 percent more. That is the honest shape of the trade: a real discount set against a real capability gap. The discount is the half that now moves twice a day.
Price the clock, not the token
Two items belong in the next vendor review. Move any workload that can wait out of 03:00 to 06:00 and 08:00 to 12:00 Central European Summer Time, because those are the expensive hours and the European afternoon is not. Then put the 24-hour notice period in front of whoever signs the budget, because a unit price that can double on a day's warning is a commercial term rather than a technical detail.
The second item is a routing decision. If a workload cannot absorb a doubled unit cost during a European morning, it needs a second provider configured, tested and priced before that morning arrives. Building that path after the notice email lands means building it inside 24 hours.
Read next: A Frontier Model Just Stopped Taking Customers | Independent Tests Rank Kimi K3 Above Fable 5
