Baidu says new AI model cuts training costs sharply

Baidu released its Ernie 5.1 model and said it delivers comparable performance to top systems at about 6% of the pre-training cost of similar models.

The company said Ernie 5.1 reduced computing needs compared with earlier versions.

On the Arena benchmark, Ernie 5.1 ranked first among Chinese models and fourth overall, behind OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.6 and 4.7.

Some AI industry figures questioned Baidu’s cost claims.

Still, the launch reflects a wider push by Chinese firms to build cheaper models under US chip restrictions.

🔗 Source: Chosun Daily

🧠 Food for thought

Implications, context, and why it matters.

Baidu’s efficiency push answers shifting US chip rules

Baidu plans as if leading-edge graphics processing units (GPUs), the chips used to train AI models, will stay out of reach. Its current stockpile should last one to two years, said Robin Li, CEO of Baidu 1.
The stance grew out of US export controls that started in October 2022. Washington tied the rules to national security concerns, including military modernization and human rights abuses 2.
US policy has also swung sharply, from a ban on AI chip exports to China in April 2025 to a reversal in July 2025 3.
By early 2026, China had approved imports of Nvidia H200 chips for ByteDance, Alibaba, and Tencent, clearing more than 400,000 units overall 3.
That uncertainty, along with smuggling cases such as a US$2.5 billion scheme that routed Nvidia chip-equipped servers to China through intermediaries, has made computing efficiency a survival tactic for Chinese AI firms 2.

Cost-efficient AI models could reshape global technology

Baidu’s strategy puts pressure on Western AI labs to defend huge computing budgets when rivals can reach similar results at far lower cost 4.
Lower running costs could help Chinese AI firms win price-sensitive work, especially in government and legal settings where operating expense shapes buying decisions 4.
Baidu built this system with a “Once-For-All” elastic training framework. The method tunes many smaller model designs in one pre-training run and lets Ernie 5.1 be pulled out as an optimized sub-network from Ernie 5.0’s sub-model matrix 5.
If these methods spread, the industry may move away from a parameter-scaling race toward AI model designs that use capital with more care 4.