---
title: "Harvey built the legal AI exam; Z.ai just passed 19% of it at about 15 cents a task"
description: "Z.ai's GLM-Legal passed 19 percent of Harvey LAB at about 15 cents a test, and Vals calls it the cheapest model in the top 10. The benchmark Harvey built to prove its product now prints the cost of competing with it."
publisher: "The Inference"
section: "Frontier"
published: 2026-10-13T15:28:40.830Z
modified: 2026-10-13T15:28:40.830Z
canonical: https://theinference.org/article/harvey-built-the-legal-ai-exam-z-ai-just-passed-19-of-it-at-about-15-cents-a-task
language: en
keywords: "Benchmarks, Legal AI, AI Pricing, Z.ai, Harvey, Model Releases"
---

# Harvey built the legal AI exam; Z.ai just passed 19% of it at about 15 cents a task

> Z.ai's GLM-Legal passed 19 percent of Harvey LAB at about 15 cents a test, and Vals calls it the cheapest model in the top 10. The benchmark Harvey built to prove its product now prints the cost of competing with it.

The biggest legal-AI news this week is a price. On October 12, Z.ai, the Chinese lab also known as Zhipu, [released GLM-Legal, a legal agent](https://z.ai/blog/glm-legal) built on its cheap GLM-5.3-Flash base [2][3]. Independent evaluator Vals AI [scored it 19 percent on Harvey LAB](https://www.vals.ai/models/zai_glm-legal), the 1,250-task exam of law-firm work that [Harvey](https://theinference.org/markets/companies/harvey), the private company selling legal AI to the largest firms, built and open-sourced [1][3]. A task passes only if every rubric criterion does, and Vals priced a full run at about 15 cents a test [1][3]. The exam built to prove Harvey's product now prints the cost of competing with it.

## The score tripled without a bigger model

The jump came from training, not scale. Z.ai put its lightweight base through legal mid-training and reinforcement learning across tens of thousands of legal task environments, re-tuning the agent harness, the runtime that manages context and tools, in lockstep with the model [2]. The pass rate nearly tripled from the base model: seventh of 79 on Vals' board, ahead of [Anthropic](https://theinference.org/markets/companies/anthropic)'s Claude Fable 5 and [OpenAI](https://theinference.org/markets/companies/openai)'s GPT-6 Astra, behind [Meta](https://theinference.org/markets/companies/meta)'s Muse Spark 1.2 at 25 percent [1][3].

## The exam became a price list

Vals runs the held-out tasks with the internet off and prints cost beside score, calling GLM-Legal the cheapest model in the top 10 [1]. A benchmark built to prove a product has become the industry's comparison layer: the same Monday, [Meta launched Muse Wonder with Harvey LAB numbers on its own charts](https://decrypt.co/380768/meta-muse-wonder-zuckerberg-frontier-ai-race), 13 percent for its largest model [5].

*Chart: **GLM-Legal sits at the floor of the Harvey LAB cost curve** Bar chart of cost per Harvey LAB test: GLM-Legal $0.15, Muse Spark 1.2 $2.09, Kimi K3 $3.83, Gemini 4 Argon $10.55, Claude Fable 5 $19.23, GPT-6 Astra $26.16.*

*Cost per Harvey LAB test as run by Vals AI, October 2026. GLM-Legal's bar is an estimate: Z.ai publishes no GLM-Legal rate card, and Vals prices it at GLM token rates. Sources: Vals AI, MarkTechPost, AInvest. [1][3][4]*

## Cheap was the easy part

That price carries an asterisk: Z.ai publishes no GLM-Legal rate card and grants access only by application, so 15 cents is Vals' arithmetic at GLM token rates [1][3][4]. Throughput is the real trade, at [72 minutes a task on average](https://www.marktechpost.com/2026/10/13/zai-glm-legal-harvey-lab-legal-ai-model/) [3]. Google's [Gemini 4 Argon finishes the same run in 44](https://www.ainvest.com/news/zhipu-glm-legal-parity-quality-floor-test-cost-curve-bought-64-slower-throughput-2610/) [4]. A one-in-five pass rate means the model still fails most of the exam it was trained for. The price is observed; work leaving lawyers' desks is projection.

## The money sits between the benchmark and the matter

Harvey's business runs the other way. It [added more than $100 million of annual recurring revenue in a single quarter](https://theinference.org/article/5x-in-19-months-inside-harveys-15-5-billion-legal-ai-machine), with growth-equity stakes from [Goldman Sachs](https://theinference.org/markets/companies/goldman-sachs) and JPMorgan following in July [6]. Its CEO [told Nikkei it opens a Japan office in early 2027](https://asia.nikkei.com/business/technology/artificial-intelligence/us-legal-ai-startup-harvey-to-open-japan-office-in-2027-ceo-says), targeting Japan's largest firms and corporate legal departments, with efficiency pressure coming for billable time [7]. Firm AI chiefs are untroubled: Herbert Smith Freehills Kramer's chief AI officer [told Law.com she would not call skyrocketing AI utilization concerning](https://www.law.com/americanlawyer/2026/10/13/ai-utilization-is-skyrocketing-but-law-firms-arent-sweating-consumption-pricing-/) [8]. If utilization keeps rising while the exam prints a cost floor, the spread between subscription revenue and inference cost has to land somewhere, and today it lands on the firms.

The bottleneck has moved off price onto reliability and workflow, and both sides are working that gap now. Z.ai is [bolting GLM-Legal into AutoClaw](https://www.unite.ai/z-ai-unveils-glm-legal-an-agentic-model-for-legal-workflows/), a desktop agent that returns contract edits as tracked changes in Word [9]; Harvey owns the workflow layer its customers already run.

![AutoClaw, Z.ai's desktop agent application, showing its setup screen and agent panel](https://storage.googleapis.com/theinference-public/img/6eb9d8eb0c7fe5a0.png)

*AutoClaw, Z.ai's desktop agent, where GLM-Legal is being wired into contract edits returned as tracked changes in Word. (Courtesy of Z.AI) (Courtesy of Z.AI)*

A 15-cent task does not make a 15-cent lawyer. Whoever sells the distance between the two first takes the seat revenue with it, and the capability just got cheap enough to learn on.

## Takeaway

Harvey LAB has become the legal AI industry's price list, and Z.ai's GLM-Legal now sits at its floor: 19 percent of the exam at about 15 cents a task, cheap through training rather than scale, but slow and still failing most of the work. The money has moved to whoever sells the workflow between a 15-cent task and a lawyer's desk.

## Sources

1. [Vals AI, "GLM Legal Benchmarks, Cost and Capabilities," 12 October 2026](https://www.vals.ai/models/zai_glm-legal)
2. [Z.ai, "GLM-Legal: Advancing Agentic Legal Intelligence," 12 October 2026](https://z.ai/blog/glm-legal)
3. [MarkTechPost, "Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB," 13 October 2026](https://www.marktechpost.com/2026/10/13/zai-glm-legal-harvey-lab-legal-ai-model/)
4. [AInvest, "Zhipu's GLM-Legal: Near-Parity Quality at the Floor of the Per-Test Cost Curve, Bought With 64% Slower Throughput," 13 October 2026](https://www.ainvest.com/news/zhipu-glm-legal-parity-quality-floor-test-cost-curve-bought-64-slower-throughput-2610/)
5. [Decrypt, "Meta's New Muse Wonder Finally Puts Zuckerberg in the Frontier AI Race," 12 October 2026](https://decrypt.co/380768/meta-muse-wonder-zuckerberg-frontier-ai-race)
6. [Benzinga via MSN, "Goldman Sachs, JPMorgan back Harvey after AI legal startup adds $100M in annual revenue in a single quarter," 28 July 2026](https://www.msn.com/en-us/money/other/quick-spark-goldman-sachs-jpmorgan-back-harvey-after-ai-legal-startup-adds-100m-in-annual-revenue-in-a-single-quarter/ar-AA28SZXJ)
7. [Nikkei Asia, "US legal AI startup Harvey to open Japan office in 2027, CEO says," 13 October 2026](https://asia.nikkei.com/business/technology/artificial-intelligence/us-legal-ai-startup-harvey-to-open-japan-office-in-2027-ceo-says)
8. [Law.com American Lawyer, "AI Utilization Is 'Skyrocketing,' but Law Firms Aren't Sweating Consumption Pricing," 13 October 2026](https://www.law.com/americanlawyer/2026/10/13/ai-utilization-is-skyrocketing-but-law-firms-arent-sweating-consumption-pricing-/)
9. [Unite.AI, "Z.ai Unveils GLM-Legal, an Agentic Model for Legal Workflows," 13 October 2026](https://www.unite.ai/z-ai-unveils-glm-legal-an-agentic-model-for-legal-workflows/)


---

Harvey built the legal AI exam; Z.ai just passed 19% of it at about 15 cents a task — The Inference. Canonical: https://theinference.org/article/harvey-built-the-legal-ai-exam-z-ai-just-passed-19-of-it-at-about-15-cents-a-task