# GLM-Legal: API Beta, Legal Benchmarks and What Is Available

Z.ai's specialist legal model has promising evaluation results, while access and real-world reliability remain separate questions.

By Paras Madan | 2026-10-12


Z.ai announced GLM-Legal on October 12, a model built on GLM-5.3-Flash for contract review, evidence analysis and compliance work. API access is application-based. Its planned integration with the AutoClaw desktop agent is still described as coming soon.

The interesting part is the focus on completing a legal assignment from source material. A useful answer has to preserve facts, follow the brief and produce something another person can review. Getting the general legal principle right is only one part of that job.

## How to get GLM-Legal API access

The current route is Z.ai's [GLM-Legal application page](https://z.ai/consultation/legal), which asks for company details and the intended use case. This is a beta-access request, rather than confirmation of a generally available self-serve API.

At publication, the announcement and access form do not establish public GLM-Legal token pricing, rate limits or a downloadable weight release. Those details matter before a team decides whether it can run a trial or deploy the model around confidential material.

Some third-party listings show specifications and describe the model as open-weight. Without a matching official release, I would not treat those fields as deployment instructions. A specialist model's relationship to GLM-5.3-Flash does not establish that it inherits the same commercial terms or API limits.

## What the legal benchmark results show

Z.ai's [technical announcement](https://z.ai/blog/glm-legal) reports a 19.2% all-pass score on Harvey's Legal Agent Benchmark, an improvement of 12.5 percentage points over the base model. It also describes strong results on LegalBench and PLawBench, the latter covering Chinese legal tasks.

The independent evaluator [Vals AI](https://www.vals.ai/models/zai_glm-legal) lists these results at the time of publication:

| Evaluation | GLM-Legal score | Listed position |
| --- | --- | --- |
| Harvey's Legal Agent Benchmark | 19.17% | 6 of 77 models |
| LegalBench | 86.23% | 14 of 150 models |
| Legal Research Bench | 41.83% | 22 of 76 models |

The different positions are useful context. A model can perform well at producing a requested work product without leading every kind of research or reasoning evaluation. Calling it the best legal model would flatten distinctions that matter to buyers.

## Why a 19.2% all-pass score matters

According to [Vals' benchmark methodology](https://www.vals.ai/benchmarks/hlab), a task passes only when every grading criterion passes. Its implementation uses two model judges and averages their task pass rates, with internet access disabled during generation.

This is stricter than counting individual requirements satisfied. An output could have sound analysis and still fail because it misses a required issue or mishandles the deliverable. Equally, failing all-pass grading does not mean every part of the work was wrong.

**A benchmark improvement is evidence for evaluation, not permission to remove legal review.** For a team considering this model, the next question is which failures remain: omitted facts, unsupported citations, wrong jurisdiction, formatting mistakes or something else. Each creates a different review burden.

## What changed from the base model

Z.ai describes two main training stages: additional legal material during mid-training, followed by reinforcement learning in realistic legal environments. It also says the model and its agent harness were improved together. The harness is the software coordinating context, tools and execution feedback around the model.

That approach suggests a useful way to evaluate domain AI. Suppose an agent reviews a supplier contract. The model needs to understand the instruction, but the surrounding system also has to give it the right version, retain relevant clauses and preserve proposed changes in the final document.

If a system loses a clause during document handling, better legal reasoning alone will not fix the output. A trial should therefore test the whole path from uploaded records to a reviewed deliverable. Comparing isolated chat answers would miss a substantial part of the intended workflow.

## Who should evaluate it first

Legal technology builders and in-house teams with repeatable document work have a clear reason to investigate. Start with a bounded task where reviewers can compare the source documents, required findings and final output. Use approved test material before considering live client records.

The unresolved question is whether specialization reduces the time spent checking the work. A detailed answer that introduces extra claims can create more review than a shorter one. Useful measurement should include correction time and missed issues, alongside completion rate.

GLM-Legal makes that tradeoff worth testing. The next evidence to watch is the beta's access terms and performance on complete workflows that practitioners can inspect, including where the model needs a person to take over.

I'm Paras. I work across AI, engineering and distribution, especially with technical products that need clear explanations.

If you're building an AI product and want help turning complex capabilities into useful technical stories and founder-led content, I'd be happy to talk.

[Book a 30-minute call](https://cal.com/paras-madan/30min)