Zhipu’s Z.ai released GLM-5.2 in late June, and within its first week on Vercel, daily token volume grew roughly 27-fold, the fastest adoption of any model Vercel tracked in all of 2026, according to reporting confirmed by CNBC in late June. That’s not a modest bump, it’s the kind of adoption curve that normally takes a genuinely disruptive product months to reach, compressed into days.
The number that matters most sits on OpenRouter, the routing platform many AI-native startups use to switch between model providers without rewriting their integration each time. GLM-5.2 alone accounts for roughly 75 percent of the market share among Chinese models routed through the platform, and Chinese models collectively have held above 30 percent of all tokens routed by US developer platforms since February, peaking as high as 46 percent, against an average of just 11 percent over the prior year. Chinese open-weight models have gone from a curiosity to a genuine plurality of certain developer traffic in a matter of months.
What GLM-5.2 actually ships is a 1-million-token context window built on a 744-billion-parameter mixture-of-experts architecture, released under a permissive MIT license and priced at roughly $1.40 per million input tokens. That combination, frontier-scale context length, a fully open license that permits commercial fine-tuning and self-hosting, and a price a fraction of the closed US alternatives, is precisely the profile that made the Vercel and OpenRouter numbers move as fast as they did.
GLM-5.2’s adoption looks structurally different from DeepSeek’s R1 moment in early 2025, which the market eventually treated as a one-off chatbot shock that faded once the initial novelty wore off. GLM-5.2 is specifically strong at agentic work, planning, coding, testing, and iterative looping, exactly the category of task enterprises are racing to automate right now, which is why this adoption curve is showing signs of sticking rather than fading the way DeepSeek’s initial spike eventually did.
The economics explain why. GLM-5.2 sits within a percentage point of Anthropic’s Opus 4.8 on a closely watched agentic benchmark at roughly a fifth of the cost. More broadly, open Chinese models are now running 60 to 90 percent cheaper than closed US frontier systems, and GLM-5.2 specifically is free to download, fine-tune, and self-host, which removes the per-token cost entirely for any team with the infrastructure to run it themselves rather than call a hosted API.
The market has priced Zhipu’s momentum even more aggressively than the usage numbers alone would suggest. The company went public on the Hong Kong Stock Exchange on January 8, 2026, becoming the world’s first publicly listed large-language-model company, pricing its IPO at HK$116.20 a share and raising roughly HK$4.17 billion at a valuation near $7 billion. By July 8, the stock had gained roughly 1,470 percent since listing, pushing Zhipu’s market capitalization past HK$1 trillion, more than $128 billion, on June 22 and to roughly $103.8 billion by early July. Before going public, Zhipu had raised about $1.5 billion in private funding from a backer list that reads like a cross-section of Chinese tech and state capital: Alibaba, Tencent, Ant Group, Meituan, Xiaomi, Hillhouse, Qiming Venture Partners, several Chinese local-government investment funds, and Saudi Arabia’s Prosperity7 Ventures. Riding that momentum, Zhipu launched a follow-on share placement in the days after GLM-5.2’s release seeking to raise up to HK$33.6 billion, roughly $4.3 billion, at as much as a 13 percent discount to market price, reportedly the largest single refinancing on the Hong Kong exchange in 2026.
The adoption has real limits. Many regulated-industry clients in the US and EU, banking and cybersecurity in particular, will not put a Chinese-origin model into their production stack at any price, regardless of benchmark performance, and enterprise migrations away from an incumbent provider take months even when the cost savings are obvious. The shift is real, but it’s currently bounded to companies and use cases willing to take on the provenance and compliance risk that comes with a Chinese model.
The competitive squeeze this puts on Anthropic and OpenAI is real and immediate. Both companies are simultaneously negotiating a voluntary pre-release review framework with the US government that Chinese labs face no equivalent of, meaning American frontier labs are absorbing a new regulatory compliance cost and an accelerating price-competition problem from the same direction at the same time, a genuinely difficult position for companies whose entire business model depends on premium pricing for frontier capability, while a rival with no such review process just became one of the most richly valued AI companies in the world purely on the strength of a free model.
This is arguably the single most directly actionable AI story of the year for a Philippine founder, not just one worth reading. Filipino startups building AI features on thin margins have no regulatory reason to avoid a Chinese open-weight model the way a Western bank does, and near-parity agentic performance at a fifth of the cost is the difference between an AI feature that’s a sustainable line item and one that quietly eats a startup’s unit economics. The real caveat is data handling: self-hosting GLM-5.2 means the security and compliance obligations that come with processing user data under the Data Privacy Act land entirely on the startup itself rather than on a vendor with its own compliance infrastructure, a tradeoff worth thinking through before chasing the cost savings.
Share this article