Z.AI Co. (2513.HK) GLM 5

Equity Research 17 August 2026 | 6:57AM HKT

Z.AI Co. (2513.HK): GLM-5.3 performance uplift brings post-training scaling at the focus; Neutral

We view the GLM-5.3 launch as another meaningful step-up in Chinese AI model Ronald Keung, CFA | progress, as Zhipu (Z.ai) attributed the performance uplift entirely from Goldman Sachs (Asia) L.L.C. post-training scaling (on GLM-5.2, with unchanged 744bn parameter size, and Allen Chang outperformed the latest DeepSeek V4 Pro of 1.6T parameter sizes across various | benchmarks). GLM-5.3 achieved material gains in coding, long-horizon task Goldman Sachs (Asia) L.L.C.

execution (with further cost efficiency per task) and cybersecurity performances, Steve Qiu | and we continue to await its next larger trillions-parameter sized model (GLM-5.5), Goldman Sachs (Asia) L.L.C. and various 2T to 5T parameters-sized models to be launched by other Chinese AI Damian Xie model companies over the next few months. | Goldman Sachs (Asia) L.L.C.

Key highlights from GLM-5.3 include: 1) further post-training scaling, through Iris Xiao | longer task environments, longer training duration and reinforcement learning Goldman Sachs (Asia) L.L.C.

framework upgrades (IndexShare, SAO and Slime); 2) leading open-source coding performance and improved token efficiency, with GLM-5.3 ranking amongst the top in open-source models (under MIT license without commercial restrictions/revenue sharing terms), across multiple coding benchmarks and approaching global frontier with more efficient cost metrics per task, and model weights scheduled to be released in 2 weeks; 3) training environments extending from coding problems to end-to-end expert workflows, enabling the model to independently progress from problem discovery and analysis to implementation and validation; and 4) emerging cybersecurity capabilities, with expanded capabilities across code review and vulnerability discovery/validation.

We also note other developments around harness-led data flywheels and model pricing power/more rational industry pricing normalization:

1) Harness products increasingly closing the loop across distribution, monetization and model training: Zhipu upgraded ZCode, its GLM-optimized coding harness product that surpassed 1mn active users, including remote control that enables users to manage tasks across devices, as well as multiple specialized agents collaboration. We note the introduction of idle-time tasks as a bright spot, which allows users to submit time-insensitive long horizon tasks for automatic execution during off-peak periods without consuming Coding Plan credits, to improve Zhipu’s inference capacity utilization while offering users more cost effective access. We view ZCode as strategically valuable beyond distribution and monetization, as growing usage generates long-horizon, real-world coding trajectories that can feed back into subsequent model post-training.

Goldman Sachs does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision. For Reg AC certification and other important disclosures, see the Disclosure Appendix, or go to Analysts employed by non-US affiliates are not registered/qualified as research analysts with FINRA in the U.S.

Goldman Sachs Z.AI Co. (2513.HK)

2) Increasing rational industry model pricing normalization alongside higher intelligence: Zhipu reopened GLM Coding Plan subscriptions with credit-based billing, where Lite plan is priced at Rmb118/mo, while offering a c.50% discount to equivalent API cost based on the company’s indicated weekly token allowance (yet above the effective cost under previous plans, which offered a deeper 70%-80% discount). Meanwhile, DeepSeek raised prices across its V4 APIs, with V4 Pro 0813 version’s peak-hour blended 1M token price reaching US$0.69, approaching GLM-5.2 level at US$0.88. We view these updates as further evidence that higher model intelligence and task-completion capabilities are increasingly supporting pricing…