
Open-Sourcing Harness on One Hand, Hiking Prices 500% on the Other: DeepSeek Shows Its Hand on the Same Day, Shifting from Selling Tokens to Selling Agent Productivity
Author: Claude, TechFlow
TechFlow Lead: Late night the day before yesterday, the V4 Pro official version was just released; the next night, DeepSeek announced a price increase plan: starting August 17, peak-valley pricing will be implemented, with a maximum increase of 500%, and some prices during peak hours reaching 12 times the current rate. On the same day, the Agent runtime framework DeepSeek Harness developer preview was open-sourced and launched. With one hand raising prices and the other providing tools, DeepSeek's calculation is clear: no longer just selling tokens, but starting to sell Agent productivity.
On the night of August 13, DeepSeek released an API price adjustment announcement: the new prices will take effect at 00:00 Beijing Time on August 17, adopting peak-valley pricing. Peak hours are daily 9:00 to 12:00 and 14:00 to 18:00, with idle period prices being half of the peak hours. At this point, it had been less than 24 hours since the V4 Pro official version went live; the "cheap window period" hinted in our report yesterday closed faster than expected.
Price Increase Details: Cache Hit Input Price Rises 500%, Peak Price Up to 12 Times Current Rate
According to Beijing Business Today and The Paper, the new prices for the flagship model V4 Pro are: during idle hours, per million tokens, cache hit input is 0.15 yuan, cache miss input is 4.5 yuan, and output is 13.5 yuan; during peak hours, these three double to 0.3 yuan, 9 yuan, and 27 yuan respectively. V4 Flash prices were adjusted synchronously; during idle hours, the three prices are 0.05 yuan, 1.5 yuan, and 4.5 yuan respectively, and during peak hours, they are 0.1 yuan, 3 yuan, and 9 yuan.
There are two metrics for the price increase, both worth remembering. Compared to current prices based on idle hours, the maximum increase is 500% (V4 Pro cache hit input rose from 0.025 yuan to 0.15 yuan); if compared based on peak hours, the cache hit input price reaches 12 times the current rate, and the output price is 4.5 times the current rate. Ordinary users can still use V4 Pro for free on the web platform and App, but those truly affected are developers and enterprise customers accessing the model via API.
Developers Outcry: "This Pro Has No Cost-Performance Ratio," But Still a Fraction Compared to Overseas
Developers on Weibo reacted directly: "It's not cheap anymore; the price has leveled with similar models," "Too much price increase during peak hours; this Pro has no cost-performance ratio." Some also turned their hopes elsewhere: calling for the new Ascend chips to be rolled out as soon as possible, giving DeepSeek a chance to reduce prices again. DeepSeek had previously promised that after the Ascend 950 super nodes are listed in bulk in the second half of the year, Pro prices would be significantly reduced; now it seems, the weight of this promise is even heavier.
However, switching the ruler to overseas markets, the numbers aren't that ugly. Even calculating based on the peak hour output price of 27 yuan, which converts to about 3.8 USD, it is still only a fraction of Claude Fable 5's output price of 50 USD. The change lies in the pricing logic: the price difference between cache hits and misses has been stretched to 30 times, and off-peak usage and caching have turned from "optimization items" into "cost lifelines." Moving batch processing tasks to night idle hours directly halves costs; this is the open card left for developers in the announcement.
Harness Launched on the Same Day: DeepSeek Doesn't Want to Just Sell Tokens Anymore
Almost simultaneously with the price increase announcement, the DeepSeek Harness developer preview opened for testing, with source code fully public on GitHub, receiving over 560 upvotes on Hacker News. The official page states the positioning very plainly: "Agent equals Model plus Harness"; the model is the soul, and Harness is the body that enables the Agent to understand context, call tools, and work continuously in a real environment.
To explain to non-engineering readers in one sentence: Harness is the Agent's "office system," responsible for assigning tasks, calling tools, logging, and approval processes. The feature of this office system is "everything is a plugin"; models, tools, sandboxes, storage, scheduling, and interfaces are all detachable and reconfigurable, and every step of operation has a traceable complete record. STAR Market Daily's interpretation pierced the commercial intent: DeepSeek is no longer purely selling tokens, but selling Agent productivity. Raising the model price first, then giving you a free framework that can truly make Agents usable; the ecosystem binding is more worth pondering than the price itself.
Industry Signal: Morgan Stanley Says Chinese Large Models Are "Ending Price Wars, Starting Intelligence Wars"
This price increase is not an isolated event. Morgan Stanley's research report "Ending Price Wars, Starting Intelligence Wars," released on August 9, shows that in the second quarter of 2026, the average API input price for Chinese large models has risen to 4.9 yuan per million tokens, and the output price has risen to 21.9 yuan, representing increases of approximately 48% and 80% respectively compared to the first quarter of 2025. Vendors such as ByteDance, Alibaba, Baidu, Tencent, and Moonshot AI are all included.
In other words, DeepSeek is just the last player to put down the "price weapon," and even after the price increase, it is highly likely to remain one of the cheapest in its class. For AI application entrepreneurs and compute chain investors, there are three real tracking points: the actual migration of developers after the August 17 effect (off-peak usage or switching to others); whether the delivery rhythm of Ascend 950 can fulfill the "price drop again" promise; and whether the Harness plugin ecosystem can replicate the diffusion speed of the DeepSeek model back then. Yesterday's "fraction" was the price for customer acquisition; today's price is business.














