Work in progress: Agents Directory has just launched. Stay tuned, more content is on the way.
GLM-5.2
Z.ai's coding-first GLM with a usable 1M-token context and High and Max effort levels. It launched on the GLM Coding Plan, with MIT open weights now public on Hugging Face.
Z.ai has now published GLM-5.2 benchmarks, but every number is vendor self-reported, not independently verified. The headline is Terminal-Bench 2.1: 82.7 on its best harness and 81.0 with Terminus-2, which Z.ai reports as competitive with Claude Opus 4.8 and GPT-5.5. GLM-5.2 is still not on the official Terminal-Bench leaderboard (the only Zhipu entry there is GLM-5.1 at 58.7 with Claude Code), so we are keeping it off our boards until a verified result lands. The full vendor-reported table is charted below. The MIT open weights are now public on Hugging Face, though the model is still not on OpenRouter and ships through the GLM Coding Plan.
From Z.ai's June 13 GLM-5.2 model card. These are vendor numbers, not independently reproduced. Terminal-Bench 2.1 'best harness' takes each model's strongest published agent, while the Terminus-2 rows fix the harness for a like-for-like comparison.
Source: Z.ai GLM-5.2 model card
- GLM Coding Plan
—
- Provider
Z.ai Context window
1MInput price
OpenOutput price
—Open weights
YesReleased
Jun 2026
