中国人工智能：GLM-5.3与DeepSeek重新定价改变能力-成本边界；上调智谱/MiniMax目标价，维持增持/中性评级

研报英文原文证据摘录

Asia Pacific Equity Research

16 August 2026

China Artificial Intelligence

GLM-5.3 and DeepSeek's repricing shift the capabilitycost frontier; raise Zhipu/MiniMax PT, remain

OW/Neutral

Following Zhipu’s GLM-5.3 launch and DeepSeek’s API repricing effective

Aug 17, we maintain OW on Zhipu and Neutral on MiniMax. We raise our

2026-27 revenue forecasts by 6-9% and 0-11%, respectively, and lift our PTs

to HK$1,800 and HK$260. GLM-5.3 raises Zhipu’s own capability, while

MiniMax benefits from a key rival becoming more expensive, and Hailuo H3

adds multimodal optionality. We prefer Zhipu as its improvement is internally

driven and more defensible. For MiniMax, the upcoming M3.1 is the key test

of whether its LLM can regain relevance on the capability-cost frontier.

• Capability and cost define the Pareto frontier; we currently place

greater value on the intelligence end. A model sits on the frontier when no

rival offers better capability at the same or lower price, or comparable

capability more cheaply. Both ends can support attractive

commercialization: frontier intelligence can command premium pricing and

unlock harder workloads, while cost-efficient models can capture large

volumes of mature intelligence. We currently favor capability because the

intelligence frontier is still moving quickly, stronger models are relatively

insulated from price changes by weaker alternatives, while mature

intelligence is increasingly supplied by more providers. Sustainable cost

leadership therefore requires structural efficiency, with DeepSeek providing

the clearest example in China.

Hong Kong

Internet

Olivia Xu AC

(86-21) 6106 6138

SAC Registration Number: S1730525060001

Alex Yao

(86 21) 6106 6505

SAC Registration Number: S1730523020001

Daniel Chen

(86-21) 6106 6205

SAC Registration Number: S1730521040001

J.P. Morgan Securities (China) Company

Limited

• GLM-5.3 moves Zhipu further up the capability curve; M3.1 is the next

test for MiniMax. GLM-5.3 uses the same base model as GLM-5.2 but

achieves meaningful coding and agentic gains through heavier post-training,

according to the company. With API economics broadly unchanged, we

expect stronger adoption and retention.…