GLM-Legal: Z.ai’s legal AI model for contract review and compliance

Vals AI ranks the model #7 of 79 on Harvey's Legal Agent Benchmark at 19.17%, the cheapest model in the top 10. It met 92.5% of individual grading criteria, up from 82.2% for the base model. The benchmark is strict. A task counts only if every criterion passes. The results are mixed. On Legal Research Bench it scored 41.83%, below the base model's 45.19%. Z.ai's launch chart also omits six higher-ranked models, led by Muse Spark 1.2 at 25.42%. Cost is the main selling point. Vals puts it at roughly 70 times cheaper per task than Gemini 4 Argon, which costs about $10.55. The trade-offs are access and speed. It runs only on Z.ai's hosted API by application, with no public weights. Z.ai has published no price. For founders building legal tools, the result suggests that domain training plus reinforcement learning on a small, cheap base model can narrow the gap with much costlier frontier systems. Even so, an absolute pass rate under 20% means human review remains essential. The gated beta and slow runs also make it hard to build on yet, so it is more useful as a cost benchmark than as a dependable supplier.