Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB

Z.ai released GLM-Legal, a version of GLM-5.3-Flash trained for agentic legal work like contract review, evidence analysis and compliance research. On Harvey's Legal Agent Benchmark, as run by Vals AI, it posts a 19.17% task pass rate versus 6.67% for its base model, ranking 7th of 79.

The cost number is what will get attention. Vals AI's runs put GLM-Legal at $0.15 per Harvey LAB test, against $19.23 for Claude Fable 5. If that kind of gap holds in practice, law firms and legal-tech companies have a strong reason to look at specialized, cheaper models instead of general frontier ones.

We like this story because it's a clean test of a big question: does domain training on a smaller base beat throwing a giant general model at the problem? Here, at least on one hard benchmark, the answer looks like yes. And importantly, an independent evaluator reproduced the lift. That's worth more than any launch chart.

Speaking of the chart: it compares GLM-Legal favorably with Kimi K3, Claude Fable 5 and GPT-6 Astra, and those figures match Vals AI. But it leaves out the six models ranked higher, led by Muse Spark 1.2 at 25.42%. The headline is true and also selective. Read it as strong mid-pack performance at a very low price, not as a new leader.

Keep the benchmark itself in perspective too. Harvey LAB uses all-pass grading, meaning a task only passes if every expert-written criterion passes. Even the best model clears roughly a quarter of tasks. That's a useful, demanding measure. It's also a reminder that no model here is close to doing legal work unsupervised.

There are real weak spots. GLM-Legal scored 41.83% on Vals AI's Legal Research Bench, below its own base model's 45.19%. Specialization traded away some research ability, and research is a big part of what lawyers actually do. Average Harvey LAB runs also took 1 hour 12 minutes, much slower than the rivals listed. For a firm, slow and cheap can be fine for batch review, not for a lawyer waiting on an answer.

Our read: this is a meaningful signal that specialized models can compete on legal agent work, and it adds pressure on frontier pricing. But it's application-only API access, with no published price and no public weights, so the cost advantage is still a projection from benchmark runs. We'll believe the economics when firms can buy it at a listed price.

GLM-Legal is built on GLM-5.3-Flash, described as a 320B-total, 18B-active mixture-of-experts model with a 1M-token context; Z.ai lists no separate parameter count for GLM-Legal. Training used legal mid-training followed by reinforcement learning across tens of thousands of legal task environments. Harvey LAB is an open-source benchmark of 1,250 tasks across 24 practice areas. All scores and costs cited come from Vals AI runs as of Oct. 12, 2026.