Zhipu AI's GLM-5.2 matches Opus 4.6 in code, stock jumps 14%

Zhipu AI's GLM-5.2 has matched Anthropic's Opus 4.6 to 4.8 in code generation despite a much smaller model size, the first global recognition of China's post-training capabilities since DeepSeek's pre-training advance 16 months ago.

"This is a genuine 'GLM Moment' — a major victory for Zhipu in reinforcement learning," CLSA analysts wrote in a report, calling it a milestone for China's AI post-training capabilities.

The model, released via API on June 17, achieves leading performance in code generation, with CLSA noting scores comparable to Opus 4.6 to 4.8. GLM-5.2 maintains pricing at about 8 yuan per million tokens for input and 28 yuan for output on contexts above 32,000 tokens. Shares of Knowledge Atlas Technology JSC Ltd., Zhipu's parent, surged as much as 40% intraday before closing 13.8% higher, extending a 170% rally since late March.

The advance positions Zhipu as the leading contender in China's AI race, with CLSA projecting the country could achieve Fable-level AI capabilities by the first quarter of 2027. JPMorgan raised its price target on the stock to 1,800 Hong Kong dollars, betting that GLM-5.2's performance will drive enterprise adoption and API revenue growth.

How GLM-5.2 Reshapes the Competitive Field

GLM-5.2's achievement is notable not just for its benchmark scores but for its efficiency. The model delivers Opus-class code generation with significantly fewer parameters, suggesting Zhipu has made advances in reinforcement learning and post-training optimization that its Chinese peers have not yet matched. This contrasts with DeepSeek's advance in January 2025, which focused on pre-training efficiency and infrastructure cost reduction.

The divergence is visible in the stock market. While Zhipu shares have surged 170% since late March, MiniMax Group Inc., another Chinese AI lab that listed in Hong Kong, has seen its stock fall about 50% over the same period. Analysts expect a long Zhipu-short MiniMax pair trade to gain momentum in early July, when a larger portion of MiniMax's stock becomes free from its IPO lockup.

Revenue Constraints and the Path Forward

CLSA noted that GLM-5.2 has strengthened the appeal of Zhipu's API services, but near-term annual recurring revenue growth will still depend on available computing power. The broker assigned a Hold rating, citing the need for more data on margin trends as memory-intensive applications with million-token contexts could drive up per-token inference costs.

The global AI community is watching closely. If Zhipu sustains its trajectory, CLSA believes China could achieve Fable-level AI — a tier of capability currently associated with frontier US labs — by the first quarter of 2027, or possibly earlier. For investors, the question is whether Zhipu can translate its technical lead into revenue growth before computing constraints or competitive responses from MiniMax and other Chinese labs narrow the gap.

Zhipu trades at a premium to its Chinese AI peers, reflecting the market's bet that GLM-5.2's technical edge will convert into enterprise contracts. JPMorgan's 1,800 Hong Kong dollar target implies roughly 40% upside from current levels, though CLSA's Hold rating suggests caution on near-term revenue visibility.

This article is for informational purposes only and does not constitute investment advice.