Baidu’s new AI actually outperforms the best models and costs 94% less to build.

Baidu, the Chinese equivalent of Google, Ernie 5.1 is officially released Late last week, the headline number was hard to ignore: the cost of training the model is roughly 94% less than similar AI systems at the same scale.

To put this in perspective, training a frontier AI model usually costs several times Millions (if not billions) of dollars in the account. Baidu, which controls more than 76% of China’s search market and trades on the Nasdaq under the name BIDU, claims to have achieved the same performance category for about two dozen.

The trick is called “Multi-Dimensional Flexible Pretraining.” Instead of building ERNIE 5.1 ​​from scratch, Baidu extracted an improved subnet from the existing ERNIE 5.0 architecture — which it launched in January 2026 — and compressed it. Total parameters decreased to about a third of the original. Active parameters (those that are actually doing work during the conversation) have been reduced by half. The result is a leaner model that inherits the knowledge base from its larger parent company without duplicating the full training bill.

In the LMArena Search Arena — a leaderboard where real users compare AI models on live web search tasks, scored by human preference —Ernie 5.1 record 1223It ranked fourth globally and first among all Chinese models. Its proxy capabilities — how well it handles multi-step tasks like filling out spreadsheets or browsing the web autonomously — have surpassed DeepSeek-V4-Pro, the previous Chinese standard.

Ernie may seem like an exotic name among Western observers, but he is actually a major role model in China. Baidu launched Ernie Bot in August 2023, and The chatbot has reached 100 million users in China By December of that year – faster than most Chinese competitors, although still slower than ChatGPT’s two-month world record to reach the same feat.

The efficiency angle mirrors what DeepSeek did to the AI ​​industry in January 2025. When the Chinese startup launched the R1, a prototype O1 matching for OpenAI with 98% lower query cost– It has wiped $600 billion off Nvidia’s market value and forced every major AI lab to rethink whether throwing computing at problems is the only viable strategy. ERNIE 5.1 ​​is a different story in terms of efficiency – in terms of training, not inference – but the basic message is the same: Chinese labs continue to find ways to do more with less.

The post-training pipeline is also worth noting. Baidu has built a four-stage reinforcement learning system called MOPD (Multi-Parameter Policy Distillation). Instead of trying to teach every skill at once—which leads to “see-through effects,” where, for example, improving math performance impairs creative writing—Baidu trained expert models in parallel on code, inference, and agent tasks, and then distilled them all into one unified model. The final online reinforcement learning phase dealt with open conversations and creative production, preserving what the distillation process could not capture well.

In theory, this should mean that all skills are equalized in terms of proficiency, without prioritizing one over the other

In GPQA (Google Graduate Level Questions and Answers, a standard for measuring whether a model can answer expert-level scientific questions that cannot be searched on Google), ERNIE 5.1 ​​comes close to the performance of leading Western closed-source models. In AIME26 – the 2026 American Invitational Mathematics Exam Modified, which tests advanced problem solving under competition conditions – the model scored 99.6% when using tool-assisted reasoning, trailing only Gemini 3.1 Pro.

Baidu says ERNIE 5.1 ​​is already being rolled out across more than 10 creative and agent platforms in China, including AI role-playing platforms and short-form drama generation tools. The form is available at ernie.baidu.com And via the Application Programming Interface (API) on Baidu’s AI Cloud platform.

Baidu is hosting its annual Create 2026 developer conference from May 13-14 in Beijing, where it plans to showcase ERNIE’s latest industrial applications. This event will serve as the next data point on how intent the company is to push the model into enterprise and global markets.