Vals Legal Index

Motivation

The Vals Index weights its sectors by their share of GDP, so work that matters more to the economy counts for more. The Vals Legal Index applies the same idea within law: practice areas are weighted by their share of U.S. law-office revenue, and each area is scored with the benchmarks that cover it. Legal Research Bench covers research, HLAB-Vals-1.1 covers drafting and reviewing documents, and Tax Agent Bench covers tax law.

Results

Practice-area scores for the top model from each of the three leading labs. The selectors swap in any other model.

The same scores on a map of the market. Each tile is a practice area sized by its share of law-office revenue, colored by the score there, averaged over every listed model or for one model you pick. The grey tile is revenue no benchmark covers.

Key Takeaways

• The top of the board is tight: the leading models sit within a few points of each other, about two standard errors, so the order at the top is not settled.
• Models are better at drafting legal documents than at answering legal research questions, so the index favors strong work product over strong research.
• Business and commercial work carries the most weight because it earns the most law-office revenue. Real estate, employment, and estates are scored on work product only; family law on research only.
• No single model leads every practice area: Wonder Full leads Business & Commercial (77.13%), Real Estate (98.40%), Labor & Employment (93.45%), and Wills, Estates & Trusts (96.56%); Gemini 4 Argon leads Other Civil (80.16%) and Criminal (75.80%); GLM 5.3 leads Civil Negligence (75.01%); Wonder, Gemini 4 Argon, and Qwen 3.8 Max 0902 tie on Family (36.36%).

Methodology

Weights

The weights come from the Census Bureau’s 2022 Economic Census, which reports law-office revenue (NAICS 541110, Offices of Lawyers) by type of legal work. Law offices reported $356.0B that year. The index uses the eight practice-area lines plus the taxation sub-line, and leaves out legal support, document filing, notarization, and mediation, which no benchmark covers.

Taxation law is a sub-line of business and commercial law in the Census. Tax Agent Bench carries it at that sub-line’s share of revenue; HLAB’s tax tasks stay in the Business & Commercial line, where the Census puts them.

Scoring

Each benchmark contributes the number its own page publishes: Legal Research Bench its all-pass accuracy by area of law, HLAB its criteria pass rate by practice area (the share of a task’s grading criteria that passed), and Tax Agent Bench its weighted pass rate, the row the Vals Index uses.

Where both Legal Research Bench and HLAB cover a line, the line is their average; where only one does, it is that benchmark’s score. Within a line, Legal Research questions pool by count and HLAB areas average, since every HLAB area has five tasks. The lines then combine by revenue:

AreaScore(i) = AVG(LegalResearch(i), HLAB(i))   # over the benchmarks that cover line i
Lines = SUM(w_i * AreaScore(i)) / SUM(w_i)        # w_i = 2022 law-office revenue of line i
Legal_Index = 0.978 * Lines + 0.022 * TaxAgentBench

A model is listed once all three components have a published result for it. Latency and cost on an index row are the mean of the three component rows; token totals are their sum. The index standard error pools the component errors with the same weights, so the widest per-area errors sit on the lines only Legal Research Bench covers (family, criminal), where strict all-pass over a few dozen labelled questions is noisy. A Legal Research question can carry more than one practice-area label, and the pooling treats the area errors as independent, so the standard errors run slightly narrow. The Economic Census runs every five years, so these are 2022 weights until the 2027 census is published.

Updates

10/23/2026

Released Vals Legal Index v1: Legal Research Bench and HLAB-Vals-1.1 scored by practice area and weighted by 2022 law-office revenue, with Tax Agent Bench as the tax-law line.