Z.ai's open-weights GLM-5.2 lands near the top of the coding leaderboards at a sixth of the cost

Z.ai released GLM-5.2 on June 16, an open-weights model the company says now sits fourth on Artificial Analysis' intelligence leaderboard and second on Code Arena's front-end coding board, at roughly one-sixth the per-token cost of the leading closed American models. The weights are public, so anyone with the hardware can download and run it. That last part is the actual story.

For two years the standard line on Chinese frontier labs was "a few months behind." GLM-5.2 does not erase that. It is the first open model whose coding and agent numbers land in the same bracket as the closed systems from OpenAI and Anthropic, instead of a tier below.

What it is

GLM-5.2 is the flagship of Z.ai, the Beijing lab formerly called Zhipu AI. It carries 750 billion total parameters and a 1 million-token context window, which is the room it has to hold a large codebase or a long agent run in working memory at once. It is built for long-horizon work: multi-step jobs and autonomous agents that grind through a problem with less hand-holding, rather than single-turn chat.

Translate the jargon once. "Long-horizon" means the model is judged on whether it can finish a task that takes many steps, like resolving a real GitHub issue end to end, not whether it answers one question well. "Open-weights" means the trained model file is downloadable. It does not mean the training data or the full recipe is public. Those are different things, and Z.ai has released the former, not the latter.

About those benchmarks

Here is where it pays to read the fine print. The leaderboard placements come from Z.ai. Artificial Analysis and Code Arena are public boards, but the claim about where GLM-5.2 sits on them was made by the company, at a Beijing press briefing.

The sharper head-to-head numbers going around in secondary coverage put GLM-5.2 ahead of GPT-5.5 on a couple of software-engineering suites:

Treat those as claims, not findings. They travel through write-ups, not an audited model card I could trace to a fixed methodology and test harness, and a two-point gap on a coding benchmark is the kind of margin that moves with prompt format, scaffolding, and how many attempts you allow. The honest version: GLM-5.2 is competitive with the top closed coding models, and it is much cheaper to run. The exact ranking is noisier than any single number makes it look.

"It's the first time that an open-source model really delivers very solid coding and agent performance that can compare with the leading proprietary AI companies like Anthropic and OpenAI," Qinkai Zheng, who leads Z.ai's CodeGeeX team, told Reuters. It is a big claim. Unlike the AGI talk further down, it is one the open weights let outsiders actually test.

The part Washington will notice

Z.ai says GLM-5.2 ships with inference support for domestic chips, including Huawei Ascend clusters. Since February the GLM-5 series has been adapted to run on Chinese silicon after U.S. export controls tightened access to Nvidia's top hardware. Zheng would not say whether GLM-5.2 was trained on domestic or foreign chips, only that the lab is working to run it efficiently across architectures.

That is the detail with the longest tail. A capable open model that runs on hardware Chinese buyers can actually get is worth more inside China than a closed model they have to rent from a U.S. provider, and the timing made the point. GLM-5.2 landed a day after Anthropic cut off worldwide access to its most advanced models, which had already left some developers and governments wary of building on infrastructure one government can switch off.

The business, and the hype

The market has noticed. Z.ai shares are up more than 2,000% since the company's Hong Kong debut in January, pushing its market value past HK$1 trillion, about $128 billion. It plans a dual listing in Shanghai, has not said how much it wants to raise, and says the proceeds will fund a push toward AGI. JPMorgan projects revenue could climb more than 534% this year, with profit not expected until 2028. The revenue base is still a small fraction of OpenAI's or Anthropic's.

The AGI line is the marketing. "Our mission is to obtain AGI, so right now our focus is on how to improve our model to achieve the upper bound of intelligence," Zheng said. A planned listing funding a march toward general intelligence is a fundraising narrative, and it reads as one. The checkable claim is the cheaper, downloadable model that runs on domestic chips. The AGI framing is the part you cannot verify.

One more sign the demand is real: Z.ai has raised prices on its frontier models several times this year, even inside China's brutal AI price war, and Zheng hinted at more increases if usage keeps climbing. Labs losing the cost argument do not get to raise prices.

What to watch

Z.ai says GLM-5.5 is due in August, aimed again at long-horizon tasks and self-evolving agents. GLM-5.2 is already enough to make the lab impossible to wave off. Whether this was one good release or the start of a durable shift is the question the next model answers.