DeepSeek has raised API prices for its models and, on the same day, opened access to DeepSeek Harness, a developer-preview Agent runtime framework that has been fully open-sourced on GitHub.
The two moves came less than 24 hours after the formal release of V4 Pro. According to the report, DeepSeek announced on the evening of Aug. 13 that its new API pricing will take effect at 00:00 Beijing time on Aug. 17. The company will adopt peak and off-peak pricing, with peak hours set at 9 a.m. to 12 p.m. and 2 p.m. to 6 p.m. each day. Off-peak prices will be half of peak rates.
Set side by side, the pricing change and the Harness rollout point to a clear shift in business focus. DeepSeek is no longer only selling tokens. It is moving toward selling Agent productivity.
New API pricing starts on Aug. 17
According to the report, DeepSeek released the pricing notice less than a day after V4 Pro officially went live.
Pricing compiled by Beijing Business Today and The Paper shows that V4 Pro will cost RMB 0.15 per 1 million tokens for cached input hits during off-peak hours, RMB 4.5 for uncached input, and RMB 13.5 for output. During peak hours, those prices double to RMB 0.3, RMB 9, and RMB 27.
V4 Flash is being repriced at the same time. Its off-peak rates will be RMB 0.05 for cached input hits, RMB 1.5 for uncached input, and RMB 4.5 for output per 1 million tokens. Peak-hour prices will be RMB 0.1, RMB 3, and RMB 9.
Largest increase is 500%, with some peak prices reaching 12x current levels
The report highlights two ways to read the increase.
Using off-peak prices versus current prices, the biggest increase is 500%. It cites V4 Pro’s cached input hit price, which rises from RMB 0.025 to RMB 0.15.
Using peak pricing as the comparison, the jump is steeper. The cached input hit price reaches 12 times the current level, while the output price rises to 4.5 times the current level.
Regular users will still be able to use V4 Pro for free on the web and in the app. The direct impact falls on developers and enterprise customers connecting to the model through the API.
Developer backlash focused on value for money
Developers on Weibo responded quickly. The report quotes reactions such as 「It’s not cheap anymore, the price is already in line with similar models」 and 「Peak-hour price hikes are too large. A Pro model priced like this has no cost-performance edge」.
Some users turned their attention to future compute supply, calling for wider deployment of Ascend chips so DeepSeek could cut prices again later. The report says DeepSeek had previously promised a large reduction in Pro pricing after Ascend 950 supernodes enter mass availability in the second half of the year. After this round of price increases, that promise carries more weight.
Still cheap against overseas peers
Even with the increase, the report argues that DeepSeek remains inexpensive compared with overseas models.
It says that even at the peak-hour output price of RMB 27, the cost works out to about $3.8, still only a small fraction of the $50 output price cited for Claude Fable 5.
The bigger change is not only the headline price. It is the pricing logic. The gap between cached and uncached input pricing has been widened to 30x, turning caching and off-peak scheduling from optimization choices into core cost controls. For batch jobs, shifting workloads to nighttime off-peak hours can cut costs in half. The report presents that as a direct signal to developers.
Harness went live the same day
At nearly the same time as the price announcement, DeepSeek opened testing for the developer preview of DeepSeek Harness. Its source code has been fully published on GitHub, and the project has received more than 560 upvotes on Hacker News, according to the report.
DeepSeek’s official description is straightforward: 「Agent = Model + Harness」. In that framing, the model supplies the core intelligence, while Harness gives an Agent the structure to understand context, call tools, and keep operating in real environments.
The report describes Harness as an office system for Agents. It handles task dispatch, tool invocation, logging, and approval flows. Its main design feature is that everything is pluggable: model, tools, sandbox, storage, scheduling, and interface components can all be swapped and recombined. Each execution step is also recorded in a complete traceable log.
The Star Market Daily, as cited in the report, interprets the release as a statement of commercial intent. DeepSeek is no longer simply selling tokens. It is selling Agent productivity. By raising model prices while offering a free framework that helps developers put Agents into real use, the company is placing more emphasis on ecosystem attachment than on price alone.
A broader shift away from price wars
The report says DeepSeek’s price increase is not an isolated event.
Morgan Stanley said in an Aug. 9 research note titled Farewell to the Price War, the Intelligence War Begins that the average API input price for large models in China rose to RMB 4.9 per 1 million tokens in the second quarter of 2026, while the average output price climbed to RMB 21.9. Those figures were up about 48% and 80%, respectively, from the first quarter of 2025. The report says ByteDance, Alibaba, Baidu, Tencent, and Moonshot AI were among the companies covered.
On that reading, DeepSeek is less an outlier than one of the last major players to put down the price weapon. Even after the increase, the report says it will probably remain one of the cheapest offerings in its class.
What the market will watch next
The report points to three follow-up issues for AI application startups and investors tracking the compute supply chain.
- How developers respond after the new pricing takes effect on Aug. 17, including whether they shift usage to off-peak periods or move to competing providers;
- Whether the delivery schedule for Ascend 950 can support DeepSeek’s earlier promise of another round of price cuts;
- Whether the plugin ecosystem around Harness can match the earlier spread of DeepSeek’s models.
The report ends with a sharp contrast: yesterday’s fractional pricing was for customer acquisition, while today’s pricing is the business.