DeepSeek Raised Prices in August, Cut Them in September: Why Your China AI Brand Score Now Depends on the Time of Day (2026)

DeepSeek has changed its API pricing three times in ten weeks, and the latest schedule charges double for queries run during Beijing working hours. For brand teams that track visibility in China's AI models, this is not a developer-side footnote. It means the model that 130 million people use every month is now being run under a cost regime that varies by hour, and any brand score measured at a single moment may describe only one slice of that regime.

What actually changed

The sequence matters, so here it is in order. In July 2026, DeepSeek introduced peak and off-peak pricing, with certain hours costing double. On August 6, it told users it would raise API prices, citing what it called an "unprecedented surge in demand" and compute constraints; per Gigazine's report, the company has resources equivalent to roughly 20,000 NVIDIA H100 GPUs, and users had experienced frequent access failures and occasionally very slow inference. No figures were published at that point.

Then, on September 9, DeepSeek announced a partial reversal effective September 10 for its V4-Flash model, as reported by 36Kr. Off-peak rates are ¥0.02 per million tokens for cache-hit input, ¥1 for cache-miss input, and ¥4 for output. Peak rates are exactly double: ¥0.04, ¥2, and ¥8. Peak hours are Monday to Friday, 9:00–12:00 and 14:00–18:00 Beijing time. Input costs returned to their pre-August levels, but output pricing remains doubled against the pre-hike baseline, and peak-hour output is four times that baseline. DeepSeek also began beta testing V4.1 Flash, pitched as cost reduction through architectural upgrades.

Why this reaches brand visibility

DeepSeek's consumer app and its API are different products, so it is worth being precise about what this pricing does and does not touch. The prices above apply to API usage. They do not directly set what the consumer app shows a user who asks for a skincare or SUV recommendation. Our analysis is about the indirect channels, and we flag where we are inferring.

First, a large share of "DeepSeek" answers that consumers see are not generated in DeepSeek's own app. DeepSeek's models are embedded in other products, including search surfaces and enterprise tools, and those builders pay per token. When output costs double, a builder has an incentive to shorten answers, cap reasoning length, or retrieve fewer documents. That is an inference, not a published finding, but it follows directly from the unit economics: output tokens are where the price stayed high, and a recommendation list with explanations is output-heavy.

Second, brand recommendations are output-heavy by nature. A cache-hit input costs ¥0.02 per million tokens at off-peak, while output costs ¥4, a 200x gap. Workloads that reuse a long cached prompt and produce a short answer are cheap; workloads that produce long comparative answers are not. For brands, this favors those that can be named in a short, decisive answer over those that need a paragraph of justification to make the shortlist. A brand whose positioning takes three sentences to explain is more exposed than one that a model can place in a single phrase.

Third, there is a measurement problem. If you audit your brand on DeepSeek once a week at 10 a.m. Beijing time on a Tuesday, you are sampling the most expensive window of the pricing schedule. If your agency runs batch checks overnight, you are sampling the cheapest. Whether answers differ between those windows is an empirical question we cannot settle from public data, but the incentive structure for variance now exists where it did not before the July change. A score reported without a timestamp is less informative than it used to be.

The audience context: DeepSeek is still a top-three surface

None of this would matter if DeepSeek were a rounding error. QuestMobile's first-half 2026 data, reported on August 4, shows Doubao at 382 million monthly active users, Qianwen at 167 million (up 5,792.9% year on year, off a small base), and DeepSeek at 130 million, placing it third among AI-native apps. QuestMobile's mid-July report put native AI apps in total at 499 million MAU, up 85.4%, and noted that competition has shifted from user acquisition to retention.

DeepSeek is therefore a smaller consumer audience than Doubao by roughly a factor of three, but it remains the model that developers and technically inclined users choose, and its models are the ones most often wrapped inside other products. That combination is exactly where API pricing leaks into what consumers see.

How the three leaders are exposed differently

The pricing story lands unevenly across the three largest surfaces. Doubao, with 382 million users, is a ByteDance consumer product whose economics are subsidized by a much larger advertising and content business, so a third-party API price move is largely irrelevant to what its users see. Qianwen sits inside Alibaba's commerce and cloud stack, where token cost is one line item among many. DeepSeek is the odd one out: it is the lean independent lab, with limited compute by its own admission, whose pricing has to carry more of its business. That is why its price changes are visible and frequent, and why a brand's exposure to cost-driven behavior is highest there.

For a marketer allocating attention, the lesson is not to move budget away from Doubao, which still dominates the audience. It is to recognize that each surface has a different failure mode. On Doubao the risk is being absent from the Douyin-linked content the model draws on. On Qianwen it is weak signals in the Alibaba ecosystem. On DeepSeek, a new risk is added: your brand's presence may vary with conditions you do not control and cannot see unless you measure across time. A single blended "China AI score" hides all three differences, which is why per-model, time-stamped tracking is the more useful unit.

What this does not tell us

Three caveats keep this honest. DeepSeek has not said that it changed how its own consumer app retrieves or ranks brands, and we have no evidence that it did. The August announcement carried no numbers, so we cannot say how large the hike was. And the September cut applies to V4-Flash specifically; other tiers may be priced differently. We are describing a change in operating conditions and what it plausibly means for measurement, not a documented change in brand ranking.

How to adapt your tracking

The practical response is cheap. Run your DeepSeek brand checks at two fixed windows, one inside Beijing peak hours (for example 10:00 on a weekday) and one off-peak (for example 22:00), and record the timestamp with every result. Do this for at least two weeks before drawing conclusions. If the answers are stable across windows, you have lost nothing and gained evidence that your score is robust. If they diverge, you have found a variable that your vendor's dashboard is silently averaging.

Also test answer length. Ask the same category question and measure whether your brand appears in the first sentence or only in a longer explanation. Brands that appear in short answers are better protected against any cost-driven shortening.

Takeaway for Brand Marketers

DeepSeek's pricing is now variable by hour, and output tokens remain twice as expensive as before August. Three actions follow. First, timestamp every DeepSeek visibility measurement and sample both peak and off-peak windows, so you know whether your score is stable. Second, write brand descriptions that compress: a one-line, unambiguous category claim is easier for a cost-constrained model to repeat than a nuanced story. Third, do not cut your DeepSeek monitoring because the audience is smaller than Doubao's; its models are the ones embedded elsewhere, and your presence there travels with them. Treat all of this as a hypothesis to test over the next month, not a settled rule.

Related: Compare how international brands score across six Chinese AI models on the GEO Hub /brands page · Earlier on the Hub: "The DeepSeek V4 Paradox: A Stronger Model, a Shrinking Audience"

Sources: 36Kr (DeepSeek API price adjustment, Sept 9, 2026); Gigazine and ProPakistani (Aug 6–7, 2026 coverage of DeepSeek's price increase); QuestMobile H1 2026 data as reported by China.com (Aug 4, 2026) and IT之家 / Wallstreetcn (July 14–15, 2026).