AnswersPublished18 sources
Baidu says its search engine can improve itself. What’s behind the claim?
At an October 30 briefing, Baidu said real searches and user feedback feed a “retrieval–feedback–retraining” loop, with multiple agents improving together and iteration efficiency rising by more than 80%. Baidu’s enhanced Search AI API was promoted at 4.9 yuan per 1,000 calls through December 31, 2026; the invitatio...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- At an October 30 briefing, Baidu said real searches and user feedback feed a “retrieval–feedback–retraining” loop, with multiple agents improving together and iteration efficiency rising by more than 80%.
- Baidu’s enhanced Search AI API was promoted at 4.9 yuan per 1,000 calls through December 31, 2026; the invitation only flagship package was 9.9 yuan per 1,000 calls.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Baidu announce at its October 30, 2026, Baidu app product briefing in Beijing about its search engine’s claimed “recursive self-evo. Article summary: At its October 30, 2026 briefing in Beijing, Baidu announced claimed “recursive self-evolution” for search, an upgraded Search AI API with promotional pricing, and consumer-app features intended to turn searches into com. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
At a product briefing in Beijing on October 30, Baidu announced an upgraded Search AI API and said its search engine had achieved “recursive self-evolution.” The company’s description: real searches and user feedback become training signals, feeding a loop designed to help multiple AI agents improve together. Baidu also presented a broader shift for its app, from answering individual questions to handling tasks and delivering results.18
1
The headline performance figures, however, remain company claims. The reports reviewed do not explain how the claimed improvement in iteration efficiency was measured, or provide independent test results.3
What Baidu means by “recursive self-evolution”
Baidu says each real search and user response can feed a continuing “retrieval–feedback–retraining” cycle. The company says multiple agents can improve in coordination without manual intervention, lifting overall iteration efficiency by more than 80%.18
That describes a process for using search feedback to refine the system; it does not, by itself, show that search results are more accurate or that users get a better experience. Reports did not specify the efficiency metric, comparison baseline or evaluation period, and no independent assessment was provided in the materials reviewed.3
18
A four-layer API—and performance claims to scrutinise
Baidu describes its upgraded Search AI API as an AI-native, four-layer system spanning content, architecture, engine and application. At the content layer, the company says it draws on more than 80 languages and a “hundred-billion-scale” index of quality content. It also claimed that data credibility improved by more than 50%, without the reports reviewed defining how “credibility” was measured.2
3
Baidu has also promoted enterprise support for high concurrency and low latency, with one report citing system-level capacity of up to a million queries per second. But the supplied reporting gives no specific latency service-level agreement or test conditions. Baidu’s cloud documentation, meanwhile, lists a default limit of 3 queries per second (QPS) for the documented search service. These are different kinds of figures: the system-level claim does not establish that an individual API account can make a million calls per second.5
39
Baidu’s documentation lists MCP, OpenAPI and agent-based access. Some reports also mention Skills, but the documentation excerpt reviewed does not list that option, so its availability and scope are unclear.39
18
There is also an unresolved reporting discrepancy around the API’s reported effect on Baidu’s Wenxin assistant: one account describes a 20% improvement in “effect,” another in “efficiency.” Without a definition or test method, it is not possible to present this as a measured 20% increase in answer quality. Reports also differ on API availability, citing 99.99% and 99.999% respectively.3
4
Promotional API prices, with a comparison caveat
Reports put the enhanced Search AI API package at 4.9 yuan per 1,000 calls from October 30 through December 31, 2026, down from a listed price of 40 yuan. The package could be bought more than once during the promotion, up to a reported cumulative limit of 1 million calls; usage beyond the promotional allocation would be billed at the list price.28
3
The flagship package was listed at 9.9 yuan per 1,000 calls, against a regular price of 72 yuan, and was available through an invitation-only channel. Excess usage would also be charged at the list price.3 Baidu’s cloud documentation lists enhanced and flagship rates of 0.04 yuan and 0.072 yuan per call, matching those respective list prices when converted to a per-1,000-call rate.
39
The enhanced promotion was compared with DeepSeek’s search API, reported at 5 yuan per 1,000 successful requests—0.1 yuan more. But the descriptions use different billing terms: Baidu’s price is stated per “call,” while DeepSeek’s is per “successful request.” The small difference is a headline unit-price comparison, not proof that the services have the same billing basis, total cost or performance.28
From search results to tasks and follow-ups
For everyday users, Baidu says it wants to move beyond “one question, one answer” toward “one task, one deliverable.” Its Wenxin assistant is described as handling complex questions through multiple searches and summaries, completing routine tasks, and—after a user confirms—tracking longer-running matters and sending alerts when something changes. Baidu says the capabilities cover more than 170 tool scenarios.1
The company also said interactive features on the search results page now cover five times as many scenarios, drawing on more than 5 billion authoritative records related to government services, public services and the central bank, across more than 20 industries. Those figures come from the briefing and have not been independently assessed in the reports reviewed.2
Baidu’s AI camera update adds simultaneous recognition and dynamic tracking of multiple objects, as well as live conversation in nine Chinese dialects, including Northeastern Chinese, Sichuanese and Cantonese.4
Taken together, the announcements point to Baidu’s ambition to make search a tool for completing a wider range of tasks, not just finding information. Whether the feedback loop delivers the performance gains Baidu claims will require clearer measurement details and independent evaluation.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Baidu says its search engine can improve itself. What’s behind the claim?"?
At an October 30 briefing, Baidu said real searches and user feedback feed a “retrieval–feedback–retraining” loop, with multiple agents improving together and iteration efficiency rising by more than 80%.
What are the key points to validate first?
At an October 30 briefing, Baidu said real searches and user feedback feed a “retrieval–feedback–retraining” loop, with multiple agents improving together and iteration efficiency rising by more than 80%. Baidu’s enhanced Search AI API was promoted at 4.9 yuan per 1,000 calls through December 31, 2026; the invitation only flagship package was 9.9 yuan per 1,000 calls.