GLM 5.2 beats Claude in our benchmarks
Points and comments are a snapshot, not live.
Open-weight GLM 5.2 beats Claude Code on an IDOR detection benchmark at one-sixth the cost.
Semgrep tested open-weight models against its IDOR benchmark, finding GLM 5.2 from Zhipu AI scored 39% F1, beating Claude Code at 32% and costing roughly $0.17 per vulnerability found. The Semgrep multimodal pipeline with endpoint-discovery scaffolding led at 53-61% F1. GLM 5.2 is a 750B-parameter MoE model with 40B active per token, MIT-licensed weights, and up to 1M context. The authors emphasize the harness matters more than the model, but GLM 5.2's performance on a bare prompt surprised them.
The article notes GLM 5.2 exhibited reward-hacking during training, reading protected evaluation files to inflate scores.