China puts America's AI industry on notice again with Ernie X1, Baidu's new model

• Chinese tech giant Baidu released new AI models Ernie X1 and Ernie 4.5.
• The company says the models rival those from OpenAI and DeepSeek in performance per cost.
• China is increasingly embracing open-source models.

Baidu, China's answer to Google, has released two new AI models.

On Saturday, Baidu released Ernie X1, a reasoning model it said "delivers performance on par with DeepSeek R1 at only half the price."

It also launched a multimodal foundation model called Ernie 4.5 that the company said "outperforms GPT-4.5 in multiple benchmarks while priced at just 1% of GPT-4.5."

Baidu said it's also making its chatbot, Ernie Bot, free to the public on April 1, ahead of schedule.

The tech giant said it will "progressively integrate" Ernie 4.5 and X1 into its product ecosystem, including Baidu Search, China's dominant search engine.

Baidu's new releases come as Silicon Valley reckons with the cost of AI models, largely spurred by the latest drops from DeepSeek, a Chinese startup launched by hedge fund High Flyer.

In December, DeepSeek released a large language model called V3, and in January, it unveiled a reasoning model called R1. The models are considered as good or better than equivalent models from OpenAI but priced "anywhere from 20-40x cheaper," according to analysis from Bernstein Research.

OpenAI vs DeepSeek vs Baidu

Tokens are the smallest unit of data an AI model processes. Companies price models according to how many input tokens a model processes and output tokens it generates.

For Ernie 4.5, Baidu said that input and output token prices start as low as 0.004 Chinese yuan per thousand input tokens and 0.016 per thousand output tokens.

BI converted those figures into US dollars to understand how chat models from OpenAI, DeepSeek, and Baidu compare against one another. While Baidu's cost claims against OpenAI's latest, "emotionally intelligent" GPT-4.5 check out, DeepSeek's V3 barely wins out in cost against Ernie 4.5.

In terms of reasoning models, conversions to USD show that Ernie X1 is the cheapest of all with prices just under 2% of OpenAI's o1.

Cost-savings aside, those who've already tried Ernie seem impressed. "Been playing around with it for hours, impressive performance," Alvin Foo, a venture partner at Zero2Launch, said in a post on X.

Baidu's latest models demonstrate not only the growing competition between the United States and China in the AI race but also the latter's growing embrace of open-source models.

"One thing we learned from DeepSeek is that open-sourcing the best models can greatly help adoption," Robin Li, Baidu's CEO, said on an earnings call in February. "When the model is open source, people naturally want to try it out of curiosity, which helps drive broader adoption."

Baidu said on X in February that the Ernie 4.5 series will be open-sourced from June 30. The company declined to comment on the X1 model.

China, which aims to become a global leader in AI by 2030, is making waves with a slate of new models and agents, including Alibaba's open-source model, QwQ-32B, and AI agent Manus, released earlier this month.

Until now, AI insiders seemed eager for the coming launch of DeepSeek's R2. But the Ernie collection may give it a run for its money.

Correction: March 17, 2025 — An earlier version of this story misstated that Baidu's models are open-sourced. The company said Ernie 4.5 series will be open-sourced in late June and did not comment on the X1 model.