GLM-5.2

Released 16 June 2026 by Z.ai, GLM-5.2 is a Mixture of Experts model of 753 billion parameters with 40 billion active, a 1M token context, and weights published under MIT license. Artificial Analysis measured it at 1524 Elo on GDPval-AA: third overall behind Claude Fable 5 (1783) and Claude Opus 4.8 (1615), level with GPT-5.5 at xhigh (1509), and first among open weights by a wide margin over MiniMax-M3 (1408). It also finished ahead of Gemini 3.5 Flash (1357). The tasks were agentic, averaging about 31 turns each across 1,999 matches. Pricing is $1.40 and $4.40 per million tokens, roughly a sixth of frontier rates.

What makes the model persistent is that it became the measuring stick for the generation after it. Z.ai kept GLM-5.2 as the base model for GLM-5.3 and stated that "Scaling post-training is all we did for GLM-5.3", so every gain reads off GLM-5.2 directly: Terminal-Bench 3.0 from 4.6 to 28.3, ExploitBench from 24.4% to 54.4%, ExploitGym from 29/39 tasks to 105/130. Anthropic's Frontier Red Team used the same baseline from the other direction, placing GLM-5.2 at zero or near zero on autonomous construction of full exploits, which is how they located the threshold GLM-5.3 crossed.

Outside the benchmark tables it serves as the open-weights default. Mistral AI made it the first third-party model hosted under its European regional controls, and Nicolas Martignole co-wrote "The Token Manifesto" with it and MiniMax-M3.