DeepSeek is abandoning flat API pricing. Starting at midnight Beijing time on August 17, 2026, its published rates split into peak and off-peak tiers, with off-peak set at exactly half the peak price and peak hours defined as 9:00 to 12:00 and 14:00 to 18:00 Beijing time. Every rate goes up. The steepest increase falls on cached input for its top model, which rises from 0.025 yuan to 0.30 yuan per million tokens at peak -- twelve times the current price.
Key facts
- New pricing takes effect 00:00 Beijing time, August 17, 2026. Peak hours are 9:00-12:00 and 14:00-18:00 Beijing time; off-peak is half the peak rate.
- V4-Pro output goes from 6 yuan to 13.5 yuan per million tokens off-peak and 27 yuan at peak. V4-Flash output goes from 2 yuan to 4.5 and 9 yuan.
- The largest multiplier is on cached input for V4-Pro: 0.025 yuan today, 0.30 yuan at peak, a twelvefold rise.
- Primary source: DeepSeek's own model and pricing page.
For two years DeepSeek's defining product feature was a number: it was the lab that made frontier-adjacent capability absurdly cheap, and the pressure it exerted on everyone else's price list was arguably its main effect on the industry. That posture is now formally over, and the shape of the replacement says more than the size of the increase.
Time-of-day pricing is not a pricing decision. It is a capacity decision. Electric utilities, toll roads, and airlines all price by the clock for the same reason: when demand at 10am exceeds what you can serve and demand at 3am does not, you either build more capacity or you make the 10am customers pay to be there. DeepSeek has picked the second option, and it is doing so weeks after publishing V4-Pro-0813, a model with a one-million-token context window and up to 384,000 tokens of output -- specifications that describe enormous per-request compute. The concurrency limits on the same page are telling: 2,500 concurrent requests on the cheaper Flash model, 500 on Pro.
The detail that should worry engineering teams is the cache. Prompt caching is the mechanism that makes agent work economically viable: an agent re-sends a huge and mostly unchanged context on every turn, and caching means the provider charges a token rate for the repeated part. DeepSeek's cached-input rate has been startlingly low, 0.025 yuan per million tokens on Pro, which is roughly a hundredth of the uncached rate. At peak, that discount narrows sharply. Uncached input rises by a factor of three at peak, but cached input rises by a factor of twelve -- so the gap the cache buys you shrinks from about 120x to about 30x. Long-running agents that hammer the same context all day are precisely the workload being repriced, and precisely the workload that runs during business hours.
Run the multipliers and the picture is consistent rather than random. Against today's rates: uncached input rises 1.5x off-peak and 3x at peak on both models. Output rises 2.25x off-peak and 4.5x at peak. Cached input rises 2.5x and 5x on Flash, and 6x and 12x on Pro. The heaviest penalties land on output tokens and on cached context -- the two things a reasoning or agent workload produces most of.
One correction is worth making, because it circulated widely. Summaries of the change described it as a "50% to 1000%" price increase. The real range on DeepSeek's own table runs from +50 percent at the low end to +1,100 percent at the high end, and more importantly it is not a flat reset -- it is a schedule, and a team that can move batch work into off-peak windows will see far smaller increases than the headline. Confusing matters further, DeepSeek's English-language pricing docs still displayed the older flat rates while the Chinese page carried the new table, which is why some coverage reported the opposite story, a price cut. When a company publishes two price lists in two languages, read the one in the company's own language.
Why it matters beyond DeepSeek's own customers: the assumption underwriting the last year of open-weight enthusiasm has been that Chinese labs would keep hosted inference near cost and let the market route around Western pricing. DeepSeek's models still carry enormous weight on aggregators -- OpenRouter's own analysis put DeepSeek at nearly 20 percent of weekly platform tokens by early June, and Chinese models passed American ones in token share around the same time. If the cheapest hosted path gets more expensive during working hours, the calculus shifts toward self-hosting the open weights, which remain freely downloadable and are entirely unaffected by this change. That is the real escape hatch, and it is the one thing DeepSeek cannot reprice.
The honest caveat: DeepSeek has published a table, not an explanation. There is no statement about GPU supply, no capacity disclosure, and no commitment about how long the schedule lasts. The page ends with the standard line that prices may change and that DeepSeek reserves the right to change them. Whether this is a durable shift to demand shaping or a temporary response to a supply crunch is not something the primary source answers.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)