Z.ai's GLM-Legal Hits 19.17% on Harvey LAB at $0.15 a Test

By James Calloway

Z.ai's GLM-Legal scored 19.17% on Harvey's Legal Agent Benchmark (LAB), nearly tripling its base model's 6.67% and beating Claude Fable 5, Kimi K3 and GPT-6 Astra on that leaderboard. The catch comes fast: the API is application-only, Z.ai has published no price, and each Harvey LAB task takes an average of 1 hour and 12 minutes to run.

GLM-Legal is Z.ai's first purpose-built legal model, aimed at law firms, in-house counsel and compliance teams. It targets contract review, evidence analysis and compliance research. The launch matters because it shows domain-specific training delivering frontier-adjacent legal agent performance at roughly 1/128th the per-test cost of Anthropic's Claude Fable 5 — $0.15 versus $19.23 per Harvey LAB run, according to Vals AI's benchmark runs as of October 12, 2026.

What is GLM-Legal?

The model is built on GLM-5.3-Flash, Z.ai's first natively multimodal GLM-5 model. The base has 320 billion total parameters with 18 billion active (a mixture-of-experts design, meaning only a fraction of parameters fire per query), a 1M-token context window and 128K maximum output. Z.ai lists no separate parameter count for GLM-Legal itself. The base model mixes sparse and linear attention (techniques that reduce the compute cost of long documents) and accepts text, images, video and files as input.

Unlike its MIT-licensed base, GLM-Legal ships with no public weights. There is no GLM-Legal repository on Hugging Face as of October 13, 2026. In plain words: the cheap open base is replaceable, but the legal model itself is Z.ai's hosted service only.

How was GLM-Legal trained?

Training ran in two stages:

• Mid-training on expanded legal corpora, which Z.ai says strengthens reasoning from facts to rules to conclusions.
• Post-training with reinforcement learning across tens of thousands of legal task environments.

Z.ai also describes the agent harness — the software scaffold that lets the model call tools and produce work products — as co-evolving with the model during post-training. Edge cases in training exposed harness bottlenecks; harness fixes then exposed the model to more realistic constraints. That is a plausible engineering account, but no external audit of the training pipeline exists. Takeaway: the claim rests on Z.ai's own description of the process.

How credible is the Harvey LAB result?

Harvey LAB is an open-source benchmark of 1,250 tasks across 24 practice areas, graded against more than 75,000 expert-written rubric criteria. A task passes only if every criterion passes — Harvey calls this all-pass grading. Vals AI runs the held-out set with two LLM judges and internet access disabled; agents get file tools plus docx, pptx and xlsx skills.

On that leaderboard, GLM-Legal scores 19.17% and ranks #7 of 79. An independent evaluator, Vals AI, reproduced the 12.5-point lift over the base model. GLM-Legal passes 92.5% of individual criteria, in line with top models that sit near 90% to 95%.

The framing needs scrutiny, though. Z.ai's launch chart places GLM-Legal ahead of Kimi K3 (12.92%), Claude Fable 5 (11.25%) and GPT-6 Astra (5.42%) — figures that match Vals AI — but the chart omits six models ranked higher, led by Muse Spark 1.2 at 25.42%. Takeaway: GLM-Legal is a top-10 model, not the leader.

Does GLM-Legal match frontier models elsewhere?

Only roughly. Z.ai claims parity on LegalBench and the Chinese-focused PLawBench without publishing numbers. The independent figures tell a more mixed story:

• LegalBench: GLM-Legal at 86.23%, within 2.33 points of leader Claude Fable 5 (88.56%) and a hair above Kimi K3 (86.21%).
• Legal Research Bench: GLM-Legal at 41.83% — below its own base model's 45.19%, a 3.36-point regression, and well behind Claude Fable 5's 49.52%.
• Duration: 1h12m average per Harvey LAB task, slower than every listed rival, including GPT-6 Astra at 24m49s.

That Legal Research Bench regression is the number Z.ai's announcement does not mention. Specialization bought agent performance and gave back ground on research tasks.

What do the numbers look like side by side?

All scores, costs and durations are from Vals AI's runs as of October 12, 2026.

	GLM-Legal	GLM-5.3-Flash (base)	Kimi K3	Claude Fable 5	GPT-6 Astra
Developer	Z.ai	Z.ai	Moonshot AI	Anthropic	OpenAI
Params (total/active)	Built on 320B/18B base	320B / 18B	2.8T / 104B	Not disclosed	Not disclosed
Context window	1M	1M	1M	1M	1M
Weights	No public weights	Open, MIT	Open, Kimi K3 License	Proprietary	Proprietary
Harvey LAB (task pass)	19.17%	6.67%	12.92%	11.25%	5.42%
LegalBench	86.23%	83.93%	86.21%	88.56%	Not listed by Vals
Legal Research Bench	41.83%	45.19%	46.15%	49.52%	39.42%
Cost per Harvey LAB test	$0.15	$0.57	$3.83	$19.23	$26.16
Avg Harvey LAB duration	1h12m	52m17s	13m58s	26m53s	24m49s
API price (in/out, per 1M tokens)	Not published	$0.15 / $0.50	$3 / $15	$10 / $50	$10 / $50
Access	API by application	API, Hugging Face	API, Hugging Face	API	API

Vals ran both GLM models at $0.075 / $0.25 per 1M tokens. The headline cost gap — 1/128th of Claude Fable 5 — is real but computed before GLM-Legal has any published commercial price.

How can developers get access?

Law firms, legal-tech companies and in-house teams can apply for API access through Z.ai's consultation page. Z.ai says GLM-Legal will come to AutoClaw, its desktop agent app, but gave no date. For teams that want to build now, the practical path is prototyping on the open GLM-5.3-Flash — available today at $0.15 / $0.50 per 1M tokens on Z.ai's API and Hugging Face — then swapping the model ID once legal access clears. That swap assumes the two models behave comparably in production, which the Legal Research Bench gap does not guarantee.

What happens next?

GLM-Legal's ranking depends on who is counting: seventh of 79 on the full Vals AI leaderboard, first only among the rivals Z.ai chose to chart. The open questions — pricing, weights, the PLawBench numbers, and whether the 1h12m runtime shrinks — will determine whether a $0.15 legal agent is a product or a benchmark result.