GLM-5.2 becomes the leading open-weight model

A week after its release, GLM-5.2 settled into position as the top-ranked open-weight model across several independent evaluations, rather than only the vendor-reported benchmarks Zhipu had cited at launch. On the LMArena leaderboard it placed #25 for text overall — behind a cluster of proprietary models and their variants — and #10 on the agent leaderboard, trailing Anthropic’s Fable and various Opus and GPT configurations. It ranked #8 on EQ-Bench’s longform creative-writing category and took second place on Vending-Bench 2, a simulated business-management benchmark, behind only Claude Opus 4.7.

Commentary was mixed on whether the ranking translated into practical advantage. Zvi Mowshowitz, tracking the results, called the model the new best open-weight option but noted a cost paradox: GLM-5.2 was considerably more expensive to run than other open models while not clearly beating cheaper closed alternatives, since it also required more output tokens per task than rivals. One commenter he quoted argued that Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5, both run at reduced reasoning effort, were “cheaper and smarter” for many uses.

The result illustrated a recurring pattern in the open-weight race through 2026: Chinese labs repeatedly took the top position on individual leaderboards within days or weeks of a release, without those wins settling the broader question of whether open models were closing the gap with closed frontier labs on cost-adjusted capability.