glm-legal benchmark results
z.ai reported key results on major legal benchmarks:
- glm-legal ranks first on harvey’s legal agent benchmark, substantially outperforming kimi k3, fable 5, and gpt-6 astra across 24 legal practice areas;
- on the same benchmark, glm-legal achieved a score 3x higher than glm-5.3-flash;
- however, glm-legal ranks third on legalbench, trailing kimi k3 by just 0.02% and slightly behind gpt-6 astra in second place and fable 5 in first;
- plawbench shows a similarly tight distribution of results across models, consistent with legalbench. glm-legal ranks second, trailing claude fable 5 by just 0.04%.
the new model significantly outperforms its predecessor and has become much more competitive. however, its advantage in legal tasks depends on the specific benchmark.
sources
- glm-legal announcement, z.ai blog – oct 12, 2026; harvey lab, legalbench and plawbench results
- plawbench paper, arxiv – the chinese legal practice benchmark: 850 questions, expert-written rubrics
