IMPLICATOR
A 33% one-day jump in an AI lab's stock usually means its newest model just beat the competition. Zhipu's didn't. The scorecard the Chinese company published this week shows its open-weight GLM-5.2 still trailing Anthropic's Claude Opus 4.8 on most coding benchmarks. Investors bid the shares up anyway, on a wager that has almost nothing to do with which model writes better code.
Key Takeaways
- GLM-5.2's benchmark scorecard, published this week, still trails Claude Opus 4.8 on most coding tests despite a 33% stock surge.
- The rally tracks the Fable 5 export ban: an MIT-licensed open model can't be revoked by a government order; a closed one can.
- GLM-5.2 beats GPT-5.5 on several long-horizon coding benchmarks at roughly one-sixth the API cost, $5.80 versus $35 per million tokens.
- The open-weight hedge has limits: cloud API use falls under China's National Intelligence Law; self-hosting needs ~800GB and eight H200 GPUs.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
The wager is about revocability. The Friday before GLM-5.2 arrived, the U.S. Commerce Department ordered Anthropic to cut foreign access to its two most powerful models, Claude Fable 5 and Mythos 5; because the company could not verify nationality in real time, it pulled both offline for every customer worldwide. Fable 5 had launched only three days earlier, on June 9. A model released under an MIT license and downloaded to a company's own servers cannot be switched off that way, and Washington's export controls have a track record of pushing the advantage toward Beijing. "Cutting-edge intelligence should not belong to only a few, nor should it be withdrawn at any time," Zhipu said, framing the launch against the week Anthropic had just had. That is what its investors repriced.
Knowledge Atlas Technology, Zhipu's Hong Kong-listed holding company, rose as much as 48% intraday on Monday before closing up about 33%, according to The Next Web. JPMorgan lifted its price target to HK$1,400 from HK$950; Bank of America opened coverage with a buy rating. The shares have risen more than tenfold since the company's January IPO. Peter Alexander of Z-Ben Advisors estimated that about 40% of U.S.-based AI engineers were born in China and that the export order now bars many of them from the systems they helped build, warning of "brain flight" toward Chinese labs.
Zhipu shipped GLM-5.2 on June 13 with no benchmark scores, no SWE-bench number and no Terminal-Bench result. The full scorecard, the MIT weights, and a standalone API did not land until this week. For three days the stock climbed on a model no independent lab had yet been able to test. The bet was on the category itself, an open model no government order can recall, with the numbers arriving only afterward.
The numbers, when they came, made a real bull case. On the benchmarks Zhipu published, GLM-5.2 beats OpenAI's GPT-5.5 on several long-horizon coding tests, 62.1 to 58.6 on SWE-bench Pro and 74.4 to 72.6 on the FrontierSWE dominance score, with some runs scored by outside evaluators rather than Zhipu itself. At 81.0 it is the first open-weight model over 80% on Terminal-Bench, and it placed first on the crowdsourced Design Arena, ahead of Anthropic's Fable 5. It is the strongest of a run of Chinese open-weight coding models this spring. The team behind the Cline coding tool called it "a frontier-level model for a fraction of the cost," and added: "Open weights is back."
The open-vs-closed AI split, decoded every morning
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
That fraction is the part that travels. Add Zhipu's $1.40 per million input tokens to its $4.40 output rate and a full GLM-5.2 round trip runs $5.80, against $35 for GPT-5.5 and $30 for Claude Opus 4.8, roughly one-sixth the cost of the OpenAI model. As the widely-followed AI account @scaling01 put it, "frontier labs are absolutely scamming you on API pricing."
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
None of that closes the gap the headline names. Claude Opus 4.8 still leads GLM-5.2 on those same coding tests, 69.2 to 62.1 on SWE-bench Pro and 26.0 to 13.0 on the multi-hour SWE-Marathon run, and Opus kept running straight through the export order that took Fable 5 down. The lab that lost a model to Washington still ships the better coder. One developer in the busy Hacker News thread, posting as LaurensBER, called GLM-5.2 "about six months behind the frontier labs. Very similar to Opus in January." On raw capability, China did not pass the frontier this week.
The open model carries its own catch. Run through Zhipu's cloud API, GLM-5.2 is subject to China's National Intelligence Law, which obliges Chinese organizations to cooperate with state intelligence work and which U.S. officials treat as a channel for government access to that data. Self-hosting avoids that, except the model's FP8 weights run to roughly 800 gigabytes and need eight H200 GPUs, an outlay most teams trialing the model this week will not make.
So the rally makes sense, even if it isn't about the benchmarks in the headline. GLM-5.2 is a strong, cheap, near-frontier coder that trails the best closed model and cannot be taken away from you. After the week Anthropic just had, investors and developers are treating that resilience as worth a premium, and the independent benchmarks due as the weights spread will test the rest.
Frequently Asked Questions
Did GLM-5.2 beat Claude Opus 4.8 on coding benchmarks?
No. On the scorecard Zhipu published this week, Opus 4.8 still leads GLM-5.2, 69.2 to 62.1 on SWE-bench Pro and 26.0 to 13.0 on the multi-hour SWE-Marathon run. GLM-5.2 does beat OpenAI's GPT-5.5 on several long-horizon coding tests and is the first open-weight model over 80% on Terminal-Bench, but Opus, which kept running through the export order, remains the stronger coder.
Why did Zhipu's stock surge 33%?
The jump followed the U.S. order that forced Anthropic to pull Claude Fable 5 and Mythos 5 offline worldwide on June 12. GLM-5.2, released under an MIT license, can't be switched off by a government directive. JPMorgan lifted its target to HK$1,400 from HK$950 and Bank of America opened coverage with a buy rating; the shares have risen more than tenfold since January's IPO.
How much does GLM-5.2 cost compared with GPT-5.5 and Claude Opus 4.8?
GLM-5.2's API runs $1.40 per million input tokens and $4.40 per million output, $5.80 combined. GPT-5.5 totals $35 and Claude Opus 4.8 totals $30 on the same basis, making GLM-5.2 roughly one-sixth the cost of the OpenAI model. A GLM Coding Plan subscription, which meters by prompt rather than token, starts near $10 to $18 a month.
Is GLM-5.2 open source, and can I self-host it?
Yes. Zhipu released the weights under an MIT license, with no regional limits or acceptable-use restrictions. Self-hosting is possible but heavy: the FP8 weights run to roughly 800 gigabytes and need about eight H200 GPUs, beyond most small teams. Running it through Zhipu's cloud API instead exposes call data to China's National Intelligence Law.
What's the catch with a Chinese open-weight model?
Two things. The benchmark lead over the American frontier isn't there yet; Opus 4.8 still scores higher on coding tests, and much of GLM-5.2's scorecard is Zhipu's own, with independent results still thin. And the data tradeoff is real: cloud API calls fall under China's National Intelligence Law, which U.S. officials treat as a government-access risk.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
