DeepSeek V4-Pro Price Cut: What 'Permanent' Meant

What changed

DeepSeek had been running a 75 percent introductory discount on API access to V4 Pro, putting the range at $0.0145 to $3.48 per million tokens depending on whether a request hit the cache and whether it was input or output. That rate had been due to lapse on May 31, 2026, reverting to a higher list price. On May 23, DeepSeek instead made the discounted range the permanent one: $0.003625 to $0.87 per million tokens. [1]

The numbers describe a spread, not a single price. Cache hits on input tokens sit at the low end, cache misses and output tokens at the high end, and the 75 percent reduction applied uniformly across that spread rather than to one figure in isolation. [1]

The model behind the price

The pricing applies to V4-Pro, the 1.6 trillion parameter model DeepSeek shipped with open weights under an MIT license about a month earlier, alongside the smaller V4-Flash. Both carry a million token context window. The May 23 move did not change the model; it changed what running it against DeepSeek's own API costs. [1]

How permanent it turned out to be

Less than three months later, DeepSeek raised those same prices by amounts the company itself put between 50 percent and more than 1,100 percent depending on the token type, effective August 16, and split billing into peak and off-peak hours for the first time. The company said the change was meant to allocate resources more reasonably and to encourage users to schedule work around actual demand, rather than describing it as a reversal of the May decision. [2]

Atlas interpretation: Calling a rate permanent is a statement about intent, not a binding commitment, and the August increase is the clearest evidence of the difference. The May cut removed an expiration date from a promotional price; it did not remove the underlying capacity constraints that a promotional price can paper over only as long as demand stays low. Once V4-Pro reached general availability and usage climbed, DeepSeek priced the busy hours the way any provider with limited GPU capacity eventually does. The word permanent described the mechanism, a rate with no scheduled end, rather than a promise the number itself would not move. [1][2]