GLM-5.2 takes the lead among open-weights models

Zhipu's GLM-5.2 is now the strongest open-weights model on Artificial Analysis's Intelligence Index, and the numbers are worth a look before you switch. On version 4.1 of the index it scores 51, eleven points above GLM-5.1 and ahead of the other open leaders: MiniMax-M3 and DeepSeek V4 Pro at 44, and Kimi K2.6 at 43. It keeps the same size as its predecessor, 744 billion total parameters with 40 billion active, but extends the context window to 1 million tokens from 200,000, and it ships under an MIT license. On GDPval-AA v2, a test meant to track economically useful work, it reaches 1524, roughly level with GPT-5.5 running at its highest reasoning setting.

The catch is cost in tokens. GLM-5.2 burns about 43,000 output tokens per index task, up from 26,000 for GLM-5.1 and well above MiniMax-M3 at 24,000 and Kimi K2.6 at 35,000. At Zhipu's own pricing of $1.40 input and $4.40 output per million tokens, that works out to roughly $0.46 per task, nearly double GLM-5.1. So the intelligence gain comes with a verbosity cost that shows up directly on the bill.

The full breakdown is on Artificial Analysis.

Why it matters

If you run an open-weights model in production, GLM-5.2 closes much of the gap to the closed frontier under a permissive license, but the token count means you should benchmark cost on your own traffic, not just the index score, before moving.