China's latest AI model claims to be even cheaper to use than DeepSeek

BEIJING — Chinese companies are making smarter artificial quality models that are progressively cheaper to use, echoing cardinal aspects of DeepSeek's market-shaking breakthrough.

Startup Z.ai, formerly known arsenic Zhipu, announced Monday that its caller GLM-4.5 AI exemplary would outgo little than DeepSeek to use. In opposition to the logic underlying existing AI models, Z.ai said its caller GLM-4.5 is built connected what's known arsenic "agentic" AI, meaning that the exemplary automatically breaks down a task into sub-tasks successful bid to implicit it much accurately.

The caller exemplary is besides unfastened sourced, meaning it is escaped for developers to download and use.

At astir fractional the size of DeepSeek's model, GLM-4.5 lone needs 8 Nvidia H20 chips to operate, Z.ai CEO Zhang Peng told CNBC connected Monday.

That's the spot Nvidia customized for China successful bid to comply with U.S. export controls. The chipmaker said this period that the U.S. volition let it to resume those China income aft a three-month pause, but it's unclear erstwhile those shipments volition begin.

Zhang said the institution doesn't request to bargain much of the chips arsenic it has capable computing powerfulness for now, but declined to stock however overmuch Z.ai spent connected grooming the AI model. Details volition beryllium released later, helium said.

Back successful January, DeepSeek had rattled planetary investors with its evident quality to defy U.S. spot restrictions and make an AI exemplary that not lone rivaled U.S.-based OpenAI's ChatGPT, but besides undercut it successful grooming and operating costs.

DeepSeek claimed grooming costs for its V3 exemplary were little than $6 million, though immoderate analysts said that fig was based connected the company's hardware walk of more than $500 cardinal implicit time.

Z.ai said that for its caller GLM-4.5 model, it would complaint 11 cents per cardinal input tokens versus 14 for DeepSeek R1; and 28 cents per cardinal output tokens versus $2.19 for DeepSeek. Tokens are a mode of measuring information for AI exemplary processing.

Earlier this month, Alibaba-backed Moonshot released Kimi K2, which claimed to bushed OpenAI's ChatGPT and Anthropic's Claude connected definite coding capabilities. Kimi K2 charges 15 cents for each 1 cardinal input tokens, and $2.50 per 1 cardinal output tokens, according to its website.

In precocious June, OpenAI named Zhipu successful a warning astir Chinese AI progress. The U.S. has besides added the startup to its entity list that restricts American companies from doing concern with it.

Z.ai launched successful 2019 and is reportedly planning an archetypal nationalist offering successful Greater China.

The startup has raised much than $1.5 cardinal from investors including Alibaba, Tencent and Qiming Venture Partners, according to PitchBook. Aramco-backed Prosperity7 Ventures arsenic good arsenic municipal funds from the cities of Hangzhou and Chengdu are besides among Z.ai's backers, the database showed.

In the past fewer weeks, respective different Chinese companies person announced new, open-source AI models. During the World AI Conference successful Shanghai, Tencent released the HunyuanWorld-1.0 exemplary for generating three-dimensional scenes for crippled development. Last week, Alibaba announced its Qwen3-Coder exemplary for penning machine code.