On 27 October we wrote that Baidu was holding its Search API at ¥36 per 1,000 calls while everyone else in China’s agent-search market cut prices. It has now stopped holding. At a Baidu App product upgrade briefing in Beijing on the afternoon of 30 October (Beijing time), Baidu said it has “comprehensively upgraded” its search AI API, with a promotional lowest price of ¥4.9 per 1,000 calls. Baidu’s own pitch is that this is cheaper than DeepSeek’s search API.
The announcement was carried almost word for word by Sina Tech, Leiphone, NetEase Tech and ifeng.
The price, and what it is being compared with
The maths:
- Baidu’s published Qianfan list prices (billing doc, still dated 16 October when we checked on 30 October) are ¥36 per 1,000 calls for Standard, ¥40 for Enhanced and ¥72 for Flagship.
- ¥4.9 per 1,000 is a cut of almost 88% on the Enhanced list price of ¥40, and it undercuts even the Standard tier’s ¥36.
- DeepSeek’s Search API is $0.75 per 1,000 calls, or ¥5 on its Chinese price list. Baidu’s ¥4.9 sits just underneath it.
The words “during the campaign” (活动期间) and “as low as” (最低) were doing a lot of work at the briefing, and Baidu has since filled in the small print on its Baidu AI Search campaign page. The ¥4.9 price applies to prepaid call bundles for the Enhanced edition (list price ¥40). The campaign runs from 30 October to 31 December, purchases are capped at a cumulative one million calls, and once a bundle is used up, calls are billed at the full list price again. A Flagship bundle is offered at ¥9.9 per 1,000 (list ¥72), but only with an invitation code. So this is a promotion with a ceiling, not a new list price.
The timing is not subtle either. DeepSeek started billing Search and Fetch calls made through third-party agent harnesses at 00:00 Beijing time on 1 November, which was 16:00 GMT on 31 October. Baidu priced itself under DeepSeek the day before that happened. It follows Tencent opening its Yuanbao Search API on the Sogou index on 21 October. China now has three big indexes chasing agent traffic on price.
The rest of the Baidu App upgrade
- “Recursive self-evolution” (递归自我进化). Baidu says search has achieved it, but published no mechanism, benchmark or evaluation. It echoes the “AI-era evolution theory” Robin Li set out in May. For now it is a slogan, not a finding.
- Wenxin Assistant moves “from one Q&A to one deliverable”. Complex queries return a finished output, and long-running tasks are tracked with proactive reminders, across 170+ scenarios.
- The AI Lens moves from “freeze-frame” to “flowing”. It does multi-object recognition and tracking, with real-time voice chat in nine dialects, including Cantonese, Sichuanese and Northeastern Mandarin.
- GUI interactive results are up fivefold on the SERP. They draw on more than 5bn data resources from government, public services and the central bank.
My take
Ignore the “self-evolution” line until Baidu shows its working. The price cut is the real news, and it matters for two reasons.
First, it confirms that Chinese search has split in two. One product is the consumer SERP, which is turning into an answer-and-deliverable surface. The other is an index rented out by the thousand calls to agents. On the second one, Baidu, Tencent/Sogou and DeepSeek are now competing on price, and the cheapest grounding source will get called most often. If you want Chinese AI agents to cite you, being retrievable through these APIs matters as much as a blue-link ranking.
Second, look at what the consumer upgrade plugs into the SERP: government, public-service and central-bank data in GUI modules. That is the same direction as Baidu’s “AI above, doctor below” health results. Authoritative structured sources sit on top, and open-web pages compete for whatever is left underneath. For brands, the measurable question is whether you appear in the deliverable at all, not where you rank beneath it.
The one-million-call cap tells you who this is aimed at: developers testing agents, not large platforms. Baidu World on 12 November is the obvious next checkpoint.