GLM-Legal: Legal Agent Model, Benchmarks and API Access
GLM-Legal is Z.ai’s specialist legal-agent model with an official beta and API application route. It is a candidate for legal teams evaluating document and agent workflows, not a verified free public chatbot. Apply-for-access status, a benchmark result and production permission are three different things.
AI Tool Finder Editorial Team · Official documentation reviewed October 13, 2026
No API calls, paid generation or hands-on product tests were performed. Public documentation and independent evaluation records do not prove our account has access.
Official product, application-gated access
Z.ai's October 12 launch article presents GLM-Legal as a legal specialist built on GLM-5.3-Flash, with document, reasoning and agent-harness work. Its official consultation form explicitly invites applications for GLM-Legal API access. This establishes a product identity and an application path; it does not establish approval for a general user.
| Question | Verified position | Before use |
|---|---|---|
| Can anyone call it now? | Official beta/API application exists; general access unverified | Obtain provider approval and the actual API contract |
| What is the official API price? | No standalone GLM-Legal general tariff independently verified | Request pricing, quotas and billing conditions |
| Can I download the weights? | Vals labels it open-weight; an official GLM-Legal weights repository and license were not verified | Verify the provider's actual repository, model files and commercial license |
| Does GLM-5.3-Flash access include it? | Not established | Do not substitute the base model's entitlement or identifier |
| Is it live in AutoClaw? | Launch article describes a forthcoming integration | Confirm your rollout rather than treating the plan as delivery |
We did not submit an application, create an account or test an API key. This page supplies an evaluation brief; it does not certify purchase or access.
Read the legal benchmark without turning it into a quote
Vals AI's Harvey Legal Agent Benchmark (Harvey LAB), updated October 12 and checked October 13, reports the following held-out-set results. The sample consists of legal work with documents and file-system tools, not ordinary chat questions.
| Model row | Task pass rate | Cost / test | Reported duration |
|---|---|---|---|
| GLM Legal | 19.17% | $0.15 | 1 hour 12 minutes |
| Gemini 4 Argon | 19.58% | $10.55 | 43 minutes 55 seconds |
| Wonder | 17.08% | $7.08 | 22 minutes 50 seconds |
The GLM-Legal row shows low evaluation cost and substantially longer duration than these two rows. That is not a production latency guarantee or a claim that it resolves your matters more reliably. $0.15 is a test-cost observation, not an official per-request or per-document price.
Vals says a task passes only when all its criteria pass, and averages task-pass results from two judges. Its model page shows 19.17% ± 3.25 for Harvey LAB and default max reasoning; it warns that benchmark-specific settings may differ. The exact GLM-Legal held-out item count was not established in the retrieved record. Harvey's full first release contains 1,250 tasks across 24 practice areas; that corpus size must not be reported as the sample size of this particular Vals run.
Vals also displays token-cost assumptions for its model evaluation. We do not present them as a verified general GLM-Legal API tariff. The benchmark reflects its harness, tools, settings and grading; changing those can change results.
Three names that must stay separate
- Harvey LAB at Vals: the legal-agent test behind the table above.
- Harvey LAB-AA: a separately reported Artificial Analysis legal evaluation; its execution and cost figures must retain that label.
- Artificial Analysis Intelligence Index: a broader composite. Our Muse Wonder general-model comparison uses that index, not the $0.15 legal-test figure.
A benchmark pass is not a legal opinion, a jurisdictional qualification or proof of regulatory compliance.
Define a bounded job for the legal team
The following are potential evaluation tasks, informed by the official positioning. They are not capabilities independently tested by AI Tool Finder.
- Contract review assistance: extract clauses, identify issues for a lawyer and draft a comparison against an approved playbook. Require exact source locations and preserve the original document.
- Case and evidence organization: build a chronology and document index. Every assertion should point to a source; gaps and conflicting evidence must remain visible.
- Initial legislation or compliance checks: identify provisions and questions requiring investigation, with jurisdiction, effective date and source authority attached.
- Drafting: prepare a first draft or a redline for professional review. Track what changed and why; do not accept silent changes to obligations.
- Legal-team agent workflows: organize approved files and produce reviewable artifacts with bounded tools and approval gates. A model's training lineage does not establish a safe deployment.
For a useful pilot, choose a representative, authorized sample and define what an acceptable artifact contains before comparing models. Record unsupported citations, omitted issues, material errors, review time and rejected outputs. This is a proposed evaluation process; no such pilot was run here.
Professional review and confidentiality come before real matters
Model-assisted review is not legal advice. A qualified legal professional should check consequential conclusions. Verify statutes, cases and quotations against authoritative sources; confirm jurisdiction, dates and whether a rule applies to the specific facts. Do not assume an answer is correct because it includes a plausible citation.
Client contracts, evidence and personal information require a separate permission and confidentiality assessment. Confirm who may upload them, provider retention and training terms, subprocessors, location of processing and deletion controls. An application form is not a data-processing agreement. Use synthetic or properly authorized material until the organization approves the real workflow.
Keep human approval before sending a client deliverable, filing a document or taking an action that affects legal rights. A lower model bill can be outweighed by correction work or a material error; measure the total cost of a professionally accepted artifact. No model use alone guarantees compliance.
View the official API application
For broader orchestration choices, browse Automation & Agents. This specialist profile does not replace legal counsel or certify the base model's API for legal deployment.
Unverified as of October 13: individual access approval, production API identifier and tariff, official downloadable weights/license, confidentiality terms for a specific client and the timing of AutoClaw availability. No client material was uploaded and no legal task was tested.