What it is
GLM-Legal is a legal-domain model from Z.ai, the Hong Kong-listed Chinese lab formerly called Zhipu AI. Z.ai announced it on 12 October 2026 as a research post. It is built on GLM-5.3-Flash, given extra mid-training on legal corpora, and then trained with reinforcement learning across what Z.ai calls tens of thousands of simulated legal environments: contract review, evidence analysis, compliance research.
It is a model, not a product. There is no lawyer-facing app, no document management connector, and no citator. As of 21 October 2026 the only way in is a gated beta: you apply for API access through a consultation form at z.ai/consultation/legal. Z.ai says GLM-Legal is coming “soon” to AutoClaw, its desktop agent, where contract edits land in Word as tracked changes and comments. That integration is announced, not shipped. Vals’ model card lists the weights as open, but the launch post names no licence and we found no GLM-Legal repository on Hugging Face on 21 October 2026. Treat it as a closed, hosted model until weights actually appear.
Why it shows up in legal-ops evaluations
One number: cost per finished task. Vals AI ran GLM-Legal on Harvey’s Legal Agent Benchmark (LAB) at launch and scored it 19.17%, 7th of 79 models, at $0.15 per task, the cheapest model in that top 10. Vals also scored it 86.24% on LegalBench and 41.83% on Legal Research Bench. Its base model scored roughly 6.7% on LAB, so the legal training is doing real work.
Read the 19% correctly. LAB counts a task only if every rubric criterion passes, so about four of every five tasks still failed somewhere. Vals has since refreshed LAB to a 120-task held-out set (HLAB-Vals-1.1, updated 21 October 2026), and GLM-Legal is not on that board yet. Its top score is Gemini 4 Argon at 27.08% for $10.55 a task. Plain GLM 5.3 sits at 14.58%, and Claude Opus 5.5 at 9.17%. The two runs are not directly comparable, so do not rank GLM-Legal against today’s leaderboard. For how these benchmarks are built, see legal AI benchmarks.
Pricing reality
Z.ai has published no GLM-Legal price. Vals ran it at $0.075 per million input tokens and $0.25 per million output tokens. That matches a promotional GLM-5.3-Flash rate, not a quoted GLM-Legal contract. Third-party routers list base GLM-5.3-Flash at $0.15 in and $0.50 out. Plan on $0.15 to $0.30 per agentic task until Z.ai quotes you. At 500 tasks a month that is $75 to $150, against roughly $5,275 for the same volume on Gemini 4 Argon at Vals’ per-task cost. Speed is the trade: Vals’ card shows an average latency near 20 minutes per legal benchmark task.
Best for
A legal engineering lead at a legal-tech vendor or a large in-house team building a high-volume, low-confidentiality pipeline: first-pass review of public filings, regulatory research across Singapore, Malaysia and Indonesia, Chinese-law work (Z.ai also reports parity with frontier models on PLawBench, a Chinese-law benchmark), or a cheap worker model under a stronger reviewer. The ROI case is volume: thousands of tasks a month where a 35x to 70x per-task cost gap pays for a human check on every output.
Not for any firm sending privileged client documents to a hosted API without a signed data processing agreement, or a US firm whose vendor-risk policy excludes entities on the US Commerce Department’s Entity List, where Zhipu has been since January 2025.
Versus the alternatives
Harvey is the segment leader by revenue and firm count. Pick it when you need a vendor accountable for legal output, lawyer onboarding and citation checking. It is reported near $1,200 per seat per month on a 20-seat minimum, and that accountability is what the price buys.
Legora is the fastest-growing entrant. Pick it when associates need a review and drafting surface they will open on their own, not an API.
Claude for Legal and OpenAI Astra for Law are the frontier-lab options. Pick them when the deciding factor is a US contracting entity, enterprise data-retention terms, or connectors into the document systems your lawyers already use. Pick GLM-Legal over both only when cost per task dominates and the documents are not confidential.
Watch-outs
- Data jurisdiction is unresolved. Z.ai’s privacy policy names a Singapore entity and says data is generally processed in Singapore, but API customers fall under a separate data processing addendum, and the policy permits transfers to group affiliates. Guard: get the addendum in writing, confirm retention and training terms for GLM-Legal beta traffic, and route only public or synthetic documents until your general counsel signs off.
- Z.ai’s terms restrict US regulated data. US users may not process HIPAA health data, GLBA financial data or export-controlled material through the service. Guard: tag matters by data class before routing and block those classes at the gateway, not by policy memo.
- The benchmark is not the job. A 19% all-pass rate means most outputs miss something, and practitioners in the HN launch thread (more than 650 points) posted examples of AI contract tools missing clauses that mattered. Guard: run 30 to 50 closed matters through GLM-Legal and your current tool, score them by hand, and keep a lawyer review step on every output.
- Beta access is not a contract. No price, no SLA, no published model ID. Guard: build behind a model-agnostic gateway such as OpenRouter so you can switch models without rewriting the pipeline.