DeepSeek's price rise reaches a rival's cloud five days late — and the discount changes its price there
Published: 8/16/2026 · Source: BytePlus ModelArk — Pricing ↗
The pricing page of BytePlus ModelArk — ByteDance's model platform — carries a notice that the billing rates for `deepseek-v4-flash-ga-260731` will be raised at 00:00 on 21 August 2026, Beijing time (18:00 on 20 August in Warsaw). The new online-inference rates are 0.44 US dollars per million input tokens, 1.32 per million output tokens and 0.014 for a cache hit.
Those are not approximate matches to DeepSeek's own new price list. They are the same three numbers. DeepSeek raised its API to exactly 0.44 / 1.32 / 0.014 on 16 August, and the platform reselling the model is following to the cent.
Two things in that notice are worth reading closely.
**The five-day gap.** Until 21 August, ModelArk still bills the old rates: 0.14 per million input tokens and 0.28 per million output. For five days the model is cheaper on a competitor's cloud than from the laboratory that trained it — output tokens are 4.7 times cheaper on ByteDance's platform than on DeepSeek's own API during peak hours. Measured against the new price, the rise on ModelArk is a factor of 3.1 on input and 4.7 on output.
**The discount changes currency.** The centrepiece of DeepSeek's new pricing is the off-peak tariff: outside the busy hours everything costs exactly half, so 0.22 and 0.66. ModelArk has those two numbers as well — but attached to batch inference, not to a clock. The discount survives the move to a reseller; the condition attached to it does not. On DeepSeek's API you buy the lower price by waiting for the right hour of the day and still getting an answer immediately. On ModelArk you buy it by giving up an immediate answer altogether, at any hour you like.
For a team choosing where to run the model, that is the whole decision, and it is not a question of price at all: both discounts cost the same 50 per cent. One is paid in scheduling, the other in latency.
The same document dates one more change: Seedance 2.5, ByteDance's video model, gains official 1080p output on 17 August 2026.