DeepSeek shipped V4-Pro, a 1.6-trillion-parameter flagship, to general availability on August 13, and at 16:00 UTC today its API stops pretending to be a commodity. A peak/off-peak schedule takes effect that lifts some output rates by more than 1,100% versus the prior flat card, and the Hangzhou lab has stopped describing this as a discount lab experiment.
At peak, V4-Pro runs $1.32 per million input tokens on a cache miss and $3.96 per million output. Off-peak, those drop to $0.66 and $1.98. The smaller V4-Flash, still in beta, sits at $0.44/$1.32 peak and $0.22/$0.66 off-peak, against a prior flat rate of $0.14 input and $0.28 output. InfoWorld’s arithmetic puts the Flash input hike at 57%–214% and output at 136%–371% depending on hour. Reuters notes V4-Pro is priced at roughly 9x Flash on input and 14x on output, the first hard split DeepSeek has drawn between a volume tier and a complexity tier.
DeepSeek’s release notes explain the schedule is designed “to allocate resources more reasonably,” which is the polite framing of capacity strain. Fortune described the “DeepSeek death zone” last month; the peak surcharge is the pricing translation of it.
The Artificial Analysis Intelligence Index now reads 53 for V4-Pro’s reasoning variant and 40 for Flash. V4-Pro speaks OpenAI’s Responses format, and the release notes call its agent capabilities in Codex-style loops “significantly enhanced.” That’s the actual news underneath the price sheet: DeepSeek is selling a frontier reasoning model into an agent market, not a chat commodity.
The competitive geometry still favors Hangzhou, but less lopsidedly. Engadget’s board has Anthropic’s Fable 5 at $50 per million output tokens, GPT-5.6 Sol at $30, Moonshot’s Kimi K3 at $15, and OpenAI’s cheaper GPT-5.6 Luna at $1.20. Flash’s edge over Luna compresses from roughly sevenfold to threefold off-peak, and just 1.4x at peak.
Sanchit Vir Gogia, chief analyst at Greyhound Research, put the sharp version bluntly: “on paper, at peak, against the right comparator, DeepSeek’s price advantage does disappear, and in places inverts.” The saving grace is architectural. DeepSeek’s roughly 98% cache-hit discount, against an industry norm near 90%, “is the mechanism that has kept its measured cost per task at about 60% below Luna, even after Luna’s cost cut,” Gogia said.
That cache discount is doing more work now than the sticker did before. It’s also happening against a fundraising backdrop that clarifies the pricing move: Reuters reported in July that DeepSeek is raising at roughly $74 billion, weeks after closing its first outside financing of about $7.4 billion in June. Frontier valuations require frontier margins. The commodity story is over because the balance sheet needs it to be.
Sources
- https://api-docs.deepseek.com/news/news260813/
- https://www.reuters.com/technology/artificial-intelligence/deepseek-formally-releases-v4-pro-model-2026-08-13/
- https://fortune.com/2026/08/13/deepseek-increases-prices-for-ai-services-by-multiple-times/
- https://www.infoworld.com/article/4209439/deepseek-raises-some-v4-prices-by-more-than-10x-as-ai-demand-strains-capacity.html
- https://www.engadget.com/2236912/deepseek-ai-models-get-four-times-pricier/