无熊者联盟：第24期人工智能大模型情绪调查报告

Knowledge Atlas(2513 HK) Hold:GLM-5.2 made leaps in long-horizon coding China ◆Context window scaled up to 1m,GLM-5.2long-horizoncoding capability matches Opus 4.7, just short of 4.8 MAINTAIN HOLD TARGET PRICE(HKD)PREVIOUS TARGET(HKD)1900.00920.00SHARE PRICE(HKD)UPSIDE/DOWNSIDE2174.00-12.6%(as of24 Jun 2026) ◆Raising earnings on continuous uplift in model capabilityand Maintain Hold;liftDCF-basedTP to HKD1,900from HKD920 Raiseestimates and TP:Wethought the USD1bn Dec2026 ARR(AnnualizedRevenue Run Rate)guidance was too stretched. We change our mind becauseof1)GLMmodelranking back to previous level(Artificial Analysis Intelligence Index[AAII]: #6at initial launchforboth5.1/5.2), 2)domestic chip compatibility forinferencingwhichpartlyresolves our concerns on chip shortage.The proposedA-share listingandpotential HK placement (Bloomberg, 24 June 2026)can alsoreduce cash flow Why not upgrade?The biggest catalysts are model launches (share price up 21%/11%/27%/33% on the next day aftertheofficiallaunch of GLM-5.0/5.1/5.1-highspeed/5.2).Also, Zhipu’s share price rallied 13% on 17 June after CSRC made remarks tosupport model companies in A-shareIPO, and rallied15% on 22 June after Zhipu co-founder commentedthatChinese models can catch up Fable 5’s level sooner than 1Q27during an interaction with Elon Musk onX.The next positivecatalyst is the launch ofGLM-5.5 (likely in August,based on average iteration pace of c2 months).Butwe remainconcerned on risks including:1) competition:Moonshot and DeepSeek raised fundsrecentlywith DeepSeek possiblylaunchingnew iteration in June(Sina news, 1 June andWSJ, 16 June), potentiallycatching up Zhipu. OpenAI isalso improvingcoding Ritchie Sun*, CFAAnalyst, Internet ResearchThe Hongkong and Shanghai Banking Corporation Limitedritchie.k.h.sun@hsbc.com.hk+852 2822 4392 Charlene Liu*Head of Internet and Gaming Research, Asia PacificThe Hongkong and Shanghai Banking CorporationLimited, Singapore Branchcharlene.r.liu@hsbc.com.sg Whyisthemarket excited aboutGLM-5.2?Zhipuofficially offersGLM-5.2in APIon17 June, after releasing ittocoding plan users on13June. GLM-5.2improveslong-horizon task capability over GLM-5.1, with context window lifting to 1m tokens. Peishan Wang*Analyst, Internet Research Compared to GLM-5.1, computational cost per token was 2.9x lower.API pricing is keptunchanged for overseas users and China userswithlonger context window.GLM-5.2 ispriced at c1.7-5xof DeepSeek v4 pro, and18-28%/15-28% of Claude Opus 4.8/GPT-5.5based on basic input/output prices.Onthe long-horizon coding benchmarks, GLM-5.2scoresin between Opus 4.7&4.8 and the highest among all open-weight models.Moreover,availability ofAscend 950supernodefrom Huawei couldbettersupport itsinferencing in 2H26, unlocking revenue upside. The Hongkong and Shanghai Banking Corporation Limited peishan.wang@hsbc.com.hk+852 3941 7008 * Employed by a non-US affiliate of HSBC Securities (USA) Inc, and isnot registered/ qualified pursuant to FINRA regulations No country for bears| The 24thedition of the EM Sentiment Survey Issuer of report:The Hongkong and ShanghaiBanking Corporation Limited Disclosures & DisclaimerThis report must be read with the disclosures and the analystcertifications in the Disclosure appendix, and with the Disclaimer, which forms part of it. View HSBC Global Investment Research at:https://www.research.hsbc.com Financials & valuation:Knowledge Atlas Source: HSBCNote:Priced at close of 24 Jun 2026 Model capabilities tracking Pricing trends Source:Company data, HSBC Exhibit8:GLM-5.2is priced at9-26%/18-52%/15-52%/29-87%ofClaude Fable 5/Claude Opus4.8/GPT-5.5/Sonnet 4.6if considering cache input, but capability is at c85%-109%level ARR trends This is for illustrationpurpose only, as we don'texpect the ramp-up trajectory Token consumption trends Exhibit12:Zhipu’s average daily token on OpenRouter saw slight decline in Mayandimprovement inJune to datedue to GLM-5.2 launch Technologicalimprovementfrom GLM-5.2 Zhipu stated that after opening up GLM-5.2 to coding plan users, developers reported ◆Strong project-level context capacity, enabling an entire codebase to be placed within asingle reasoning workflow◆More stable long-horizon task execution, allowing complex tasks to progress continuously There are still plenty ofupside for LLM to improve its Technological improvement fromZCode (coding harness) Zhipu has optimized ZCode (coding harness) to support GLM-5.2 and encouraged coding planusers to try it.ZCode can improve users’ agentic coding experience of GLM-5.2 acrossdebugging, refactoring, project understandingetc. In addition,Zhipucancollect more high A-share listingand H-share placement Recent remarks fromthe CSRC at the Lujiazui Forum(17 June 2026)indicate that two STARMarket reform measures will be rolled out in short order. First, the regulator plans to expand theapplicability of the “Fifth Set” listing standards to the artificial intelligence sector, with a clearint