Author: Claude, Deep Tide TechFlow

Deep Tide Guide: The day before, the official version of V4 Pro was just released at midnight, and the next evening, DeepSeek announced the price increase plan: starting from August 17, peak-valley pricing will be implemented, with the maximum increase of 500%, and some prices during peak hours reaching 12 times the current price. On the same day, the developer preview version of the Agent operating framework DeepSeek Harness was open-sourced. One hand raises prices, the other provides tools. DeepSeek's calculations are clear: no longer just selling tokens, but starting to sell Agent productivity.

On the evening of August 13, DeepSeek released an API price adjustment announcement: The new prices will take effect at midnight Beijing time on August 17, adopting peak-valley pricing, with peak hours from 9 AM to 12 PM and 2 PM to 6 PM, and off-peak prices being half of the peak prices. At this point, it was less than 24 hours since the official launch of V4 Pro, and the "cheap window period" we mentioned in yesterday’s report closed faster than expected.

Price Increase Details: Cache Hit Input Price Rises by 500%, Peak Price Reaches Up to 12 Times Current Price

According to the Beijing Business Daily and The Paper, the new prices for the flagship model V4 Pro are: 0.15 yuan per million tokens for cache hit input during off-peak hours, 4.5 yuan for miss input, and 13.5 yuan for output; during peak hours, all three prices double to 0.3 yuan, 9 yuan, and 27 yuan. V4 Flash has also been adjusted with off-peak prices of 0.05 yuan, 1.5 yuan, 4.5 yuan, and peak prices of 0.1 yuan, 3 yuan, and 9 yuan.

There are two bases for the increase, both worthy of note. Based on off-peak hour comparisons with the current price, the maximum increase is 500% (V4 Pro cache hit input rises from 0.025 yuan to 0.15 yuan); if compared with peak hours, the cache hit input price reaches 12 times the current price, and the output price is 4.5 times the current price. Ordinary users can still use V4 Pro for free on the web and App, and the ones truly affected are developers and corporate clients accessing the model via API.

Developers React: "Such a Pro Has No Cost-Effectiveness," But Still a Fraction Compared to Overseas

Developers on Weibo reacted directly: "Not cheap anymore, prices are already on par with similar models", "The peak price increase is too much, such a Pro has no cost-effectiveness." Some have placed their hopes elsewhere: calling for the rapid rollout of Ascend's new chip, giving DeepSeek the chance to reduce prices again. DeepSeek had previously promised that after the Ascend 950 super nodes are mass-produced in the second half of the year, the Pro price would be drastically reduced. Now, this promise carries more weight.

However, when comparing to overseas, the numbers don't look so bad. Even at the peak output price of 27 yuan, approximately 3.8 USD, it is still just 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 pulled to 30 times, turning off-peak and caching from "optimization items" into "cost bottlenecks". Shifting batch processing tasks to nighttime off-peak periods halves costs directly—this is the straightforward clue left for developers in the announcement.

Harness Launched the Same Day: DeepSeek No Longer Wants to Just Sell Tokens

Almost simultaneously with the price increase announcement, the DeepSeek Harness developer preview was opened for testing, with the source code fully disclosed on GitHub, receiving over 560 upvotes on Hacker News. The official page clearly states the positioning: "Agent equals Model plus Harness," the model is the soul, and the Harness is the body that allows the Agent to understand context, call tools, and work continuously in a real environment.

To explain to non-engineer readers in one sentence: Harness is the "office system" for Agents, responsible for task dispatch, tool adjustment, logging, and approval processes. The characteristic of this office system is "everything is a plugin," all models, tools, sandboxes, storage, scheduling, and interfaces can be swapped and reorganized, and each step of operation has a complete traceable record. The interpretation by the Sci-Tech Innovation Board Daily reveals the commercial intent: DeepSeek is no longer simply selling tokens but selling Agent productivity. First, raise the model price, then provide you with a free framework that can truly make use of Agents—ecological binding is more intriguing than the price itself.

Industry Signal: Morgan Stanley Claims Chinese Large Models Are "Saying Goodbye to Price Wars, Launching Intelligence Wars"

This price increase is not an isolated event. Morgan Stanley's research report released on August 9, "Saying Goodbye to Price Wars, Launching Intelligence Wars," shows that by the second quarter of 2026, the average API input price of Chinese large models has risen to 4.9 yuan per million tokens, and the output price has risen to 21.9 yuan, an increase of approximately 48% and 80% from the first quarter of 2025, respectively. Companies such as ByteDance, Alibaba, Baidu, Tencent, and Moon's Dark Side are all included.

In other words, DeepSeek is merely the last player to drop the "price weapon," and even after the price increase, it is still likely one of the cheapest in its class. For AI application entrepreneurs and computing chain investors, there are three real tracking points: the actual migration of developers after the implementation on August 17 (whether to shift or switch to competitors); whether the rhythm of the Ascend 950 delivery can fulfill the "further price reduction"; and whether the Harness plugin ecosystem can replicate the rapid spread of the DeepSeek model back in the day. Yesterday's "fraction" was the customer acquisition price, while today's price is the business.