Baidu said at a Beijing briefing on October 30 that its search engine now retrains itself from live use. It calls this recursive self-evolution. Each real search, and the user feedback that follows, becomes a training signal, so models keep updating without manual tuning. Baidu puts the gain in iteration efficiency at more than 80 percent. It also launched an upgraded search AI API at a promotional price it says undercuts DeepSeek's.
Baidu AI Cloud's campaign page lists the enhanced search package at 4.9 yuan per 1,000 calls. The offer began October 30 and ends December 31; the list price is 40 yuan. Buyers can keep buying until they reach 1 million calls in total, then pay list price. A flagship package needs an invitation code and costs 9.9 yuan per 1,000 calls, instead of 72. Baidu published no independent test of its self-improvement claim.
A temporary price and a quality claim only Baidu has measured leave the proof to the buyer. The loop is plain: searches and reactions feed retraining. That does not show results improve for your queries. Buyers could ask four things: what the price is after December 31, what happens past 1 million calls, whether Baidu will test on their own queries, and how user feedback is used in training.