The company that made AI cheap is putting the price back up
DeepSeek is raising its API prices from 17th Aug’26. Depending on the model, the token type and the time of day, the increases run from roughly 50% to about 1,100% over current rates. It is also introducing peak and off-peak pricing for the first time: peak is 9am–noon and 2pm–6pm Beijing time, and off-peak sits at half the peak rate. The trigger is capacity, not strategy. DeepSeek released V4 Flash hours after OpenAI cut GPT-5.6 Luna prices by up to 80% on 30th Jul’26. Demand did what you'd expect: V4 Flash processed 7.22 trillion tokens in a single week on OpenRouter, taking the top spot. Founder Liang Wenfeng has said the company runs on roughly 20,000 GPUs in H100-equivalent terms and that the biggest gap with the US is resources, not talent.
What it actually means
If you routed your batch work, your high-volume agent loops to the cheapest token which a lot of teams did, because it was the rational call, then your unit economics change on Monday. Treat time of day as a lever now that someone's pricing it and batch overnight work has just been handed a genuine discount. So, the jobs that don't need to run at 11am shouldn't. The detail most of the coverage will skip: this is a repricing of DeepSeek's hosted service, not a change of heart about open weights. The models are still MIT-licensed and still free to download and run yourself. Which means for the first time the build-versus-rent question has a real number attached to it.
The cheapest token on the board is a subsidy, not a strategy.
1,100%top of DeepSeek's price increase range
Source: TechNode · Saturday 15 August 2026