Getty Images/iStockphoto
Baidu’s low-priced new models bring questions about cost
The Chinese tech vendor said its reasoning model matches DeepSeek’s performance, at half the price. It’s unclear how much the vendor spent to build it.
China tech giant Baidu introduced its latest foundation model and a new reasoning model, which it said can compete with Chinese startup sensation DeepSeek’s low-priced model at an even lower price.
Amid a surge of Chinese tech vendors releasing generative AI models that compete on a strong footing with models from top U.S. and European vendors, Baidu on March 16 unveiled Ernie X1, a deep reasoning model with multimodal capabilities, and Ernie 4-5, its latest foundation model, which has new multimodal capabilities.
Ernie X1 is comparable in performance to DeepSeek's R1 reasoning model, according to Baidu.
The model combines accuracy, creativity and literary talent, Baidu said, and performs well in daily conversation, logical reasoning and complex calculations.
Ernie X1’s input and output prices range from $0.28 to $1.1 per million tokens, and Ernie 4.5’s input and output prices range from $0.55 to $2.2 per million tokens. Ernie 4.5 is available now using APIs on the Baidu AI Cloud MaaS (mobility as a service) platform Qianfan. Ernie X1 will soon be available on the platform.
Comparatively, DeepSeek-Reasoner, which is based on its R1 model, costs $0.55 per million input tokens and $2.19 per million output tokens. OpenAI’s o1 reasoning model costs $15 per million input tokens and $60 per million output tokens.
The introduction of Ernie 4.5 and Ernie X1 is the latest example of Chinese vendors' growing strength in the worldwide AI market. While Chinese vendors have long competed with U.S. vendors, the arrival of AI startup DeepSeek ratcheted up the competition that has seen the release of new models from Alibaba and the emergence of AI startup Manus, which last week claimed it had released the first fully autonomous AI agent.
The multimodal capabilities of Ernie X1 and Ernie 4.5 are not unique. Top AI vendors such as Google and OpenAI have been offering multimodality for several years.
Baidu’s costs
Baidu's performance capabilities will also be difficult to confirm and can only be validated after users test them out for themselves, said Chirag Shah, a professor in the Information School at the University of Washington.
What might catch the attention of enterprises is Baidu's statement that Ernie X1's performance is on par with DeepSeek at dramatically lower prices.
"Assuming that the performance claims hold up, the real win here is the cost," Shah said. He added that the only way Baidu could offer a reasoning model for half the price of DeepSeek’s is if it had been able to lower the costs to make it.
The AI vendor then probably took the same approach as DeepSeek: fine-tuning an open source model, which significantly lowers costs, Shah said.
While Baidu plans to make the Ernie 4.5 model open source later this year, it has not provided details about how it is building it.
Moreover, prices don’t necessarily reflect the full cost of building the model, said Futurum Group analyst David Nicholson.
"It's impossible to gauge the true cost of any of these models today," Nicholson said.
Meanwhile, it's likely not a coincidence that Baidu's release of the two models comes two days before AI hardware and software giant Nvidia kicks off its annual GTC conference, he said.
"This is a global battle underway for AI dominance, Nicholson said.
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.
