Z.ai released GLM-Legal this week, a legal-specialised model built on GLM-5.3-Flash and trained for agentic legal work — contract review, evidence analysis, and compliance research. The headline number is a 19.17% all-pass rate on Harvey’s Legal Agent Benchmark (LAB), nearly 3x its base model’s 6.67%, and it lands ahead of Claude Fable 5 (11.25%) and Kimi K3 (12.92%) on that test [E4 MarkTechPost / Vals AI Harvey LAB, Oct 12 2026].
Why the number matters
Harvey LAB is 1,250 tasks across 24 practice areas graded by more than 75,000 expert rubric criteria; a task passes only if every criterion passes. GLM-Legal hits 19.17% and ranks #7 of 79 on Vals AI, behind six Meta and Google models [E3 Z.ai tech blog]. On LegalBench it scores 86.23%, close to Claude Fable 5’s 88.56% [E4 Vals AI].
The catch
- Access is application-only. There is no public API price and no released weights yet [E3 Z.ai].
- It is slow: about 1 hour 12 minutes per Harvey LAB task, versus 27 minutes for Claude Fable 5.
- Cost per test is about $0.15 — roughly 1/128 of Claude Fable 5’s $19.23 [E4 Vals AI].
- Legal Research Bench actually dropped 3.36 points versus its base model, so it is not a clean win everywhere.
Who should care
For law firms and compliance teams that can tolerate gated beta and slow runs, GLM-Legal is the cheapest path to frontier-class legal-agent scores we have seen this year. For anyone needing a production API today, Claude Fable 5 or Kimi K3 remain the practical choices. We have not run the model ourselves — this rests on Z.ai’s launch and Vals AI’s independent reproduction.
GLM-Legal is a single product we covered from this week’s monitor feed (MarkTechPost + Vals AI). We did not receive API access and have not benchmarked it; scores reflect Z.ai’s published claims and Vals AI’s independent runs.
*Affiliate disclosure: Some links on this page are affiliate links. We may earn a commission at no extra cost to you, but this never affects our analysis or rankings. See our full disclosure.*
