This article is part of our AI Agent Guides. See how models compare this week on the AI Agent Leaderboard, a ranking of rankings refreshed every week.
The Vals Index is a single score from Vals AI that measures AI models on agent work in finance, coding, legal and tax. Each sector is weighted by its share of US GDP. It tries to answer a business question: how much economically useful work can this model do?
It is one of the more business-friendly boards, since it reports accuracy next to cost and time. But its weighting is a judgment call, and you may not want it for your own decision.
Key Takeaways
- The Vals Index combines eight benchmarks across four sectors into one percentage. It was at version 2.2 as of November 2026.
- Weights follow US GDP share, so finance counts far more than tax.
- Most of its benchmarks use private test sets, which reduces the risk that models trained on the answers.
- It reports cost, tokens and time per test, which helps you weigh a small accuracy gap.
- For current positions, see the AI Agent Leaderboard or the knowledge work board. We do not state ranks here.
How the score is built
Each sector is the average of its benchmarks, and then the sectors are combined. This comes straight from the Vals Index page, checked on 2 November 2026.
| Sector | GDP weight | Benchmarks included | What they test |
|---|---|---|---|
| Finance | about 8.0 | Finance Agent v2.1, Excel Modeling Benchmark v1.1 | Financial reasoning, building spreadsheet models |
| Coding | about 5.6 | Terminal-Bench 4.0, Vibe Code Bench v1.1, Code Migration v1.1 | Command-line work, building apps, porting code |
| Legal | about 1.2 | Legal Research Bench, HLAB-Vals-1.1 | Case research, long legal work products |
| Tax | about 0.5 | Tax Agent Bench | Research-grade US corporate tax questions |
The formula is simple: multiply each sector by its weight, add them up, and divide by 15.3. Vals uses Bureau of Economic Analysis data for the shares. It openly calls this "a vast oversimplification" of how AI might affect the economy, but a useful proxy.
Here is what the weights mean in practice. By our arithmetic from the published weights, finance makes up roughly 52% of the index and coding about 37%. Legal is about 8% and tax about 3%, so a large gain in tax barely moves the total.
Why the private test sets matter
A major problem in AI testing is leakage, where test questions end up in training data. OpenAI's 2026 audit of SWE-bench Verified shows how serious it can be.
Vals says its proprietary benchmarks publish results only from a test set that stays private at all times. That is a meaningful defense.
Its setup typically has three layers:
- Public validation set: open, so you can see the type of work tested.
- Private validation set: licensed to companies for their own checks.
- Test set: always private, and the only one used for published results.
The index page says Vals aggregates six private and two public benchmarks. For the private ones, you cannot inspect every task the way you can with Terminal-Bench. You are trusting the publisher's process.
How results get graded and reported
Vals' methodology describes both strict accuracy checks and rubric-based AI judges. AI judges have known biases: a 2026 Arena study found they often favor their own answers.
For selected public benchmarks, such as Terminal-Bench, it runs every model itself under the same settings. Repeat-run benchmarks currently use three runs.
Vals calls a model through a fixed harness it controls so that differences reflect the model. Many of its benchmarks also run models inside coding and terminal agents, including a model's own native agent where one exists. Vals says it also evaluates finished products, such as legal research tools, against private task sets.
It reports more than accuracy. Each row on the index shows cost per test, tokens per test and time per test. It also charts accuracy against reasoning effort, which shows how much of a score comes from paying for extra thinking.
A hypothetical way to read a row
| Model | Index | Cost per test | Time per test | What it suggests |
|---|---|---|---|---|
| Model A | Slightly higher | Three times more | Twice as long | Pay more for a small edge |
| Model B | Slightly lower | Far cheaper | Faster | Often the better business choice |
These rows are invented to show the reasoning, not real results. For the real ones, see the Vals page. Note that accuracy on an index is not your error rate on your own work.
Why you cannot compare old and new versions
The index changes often, and it records each change on its update log. In 2026 alone it has passed through several versions.
| Date | Change |
|---|---|
| 4 May 2026 | Added Vibe Code Bench, removed the saturated CaseLaw legal benchmark |
| 13 Aug 2026 | Version 2: new finance, coding and legal benchmarks, removed SWE-bench Verified as saturated |
| 25 Sep 2026 | Version 2.1: added Tax, moved to Terminal-Bench 4.0, which lowered scores |
| 19 Oct 2026 | Version 2.2: updated Finance Agent, EMB, Code Migration, HLAB and tax scoring |
So a score from August is not comparable with one from November. Vals itself says Terminal-Bench 4.0 is substantially harder, so index scores dropped.
Infrastructure matters too, as Anthropic's tests of agent benchmarks show. If a vendor quotes a Vals number, ask for the version and the date.
Also note one small quirk. The index does not use the full Code Migration benchmark. It uses a fixed subset of 60 tasks, chosen because it closely reproduces the full result.
What it does not tell you
- Your mix may differ. A law firm cares about legal, which is only a small slice of this index. Use the sector columns, not the total.
- Scope is limited. Four sectors are not the whole economy. Healthcare and customer service, for example, are not in this formula.
- Accuracy is not sign-off. Legal, tax and finance results are tests, not advice. A qualified professional must still review real work.
- US weighting. GDP weights are based on the US economy, so they may not fit a business elsewhere.
- The publisher is a business. Vals licenses private validation sets to companies, so weigh its interests as you would any vendor's.
Vals runs narrower indexes too, including finance and legal ones listed on the same page. If you work in one sector, those are often a better fit than the headline number.
How we use it
The Vals Index is one of three boards in our overall ranking. The index contains Terminal-Bench, which is also one of our inputs, so that test partly counts twice. Keep that in mind when you read our methodology.
Our knowledge work board goes deeper with Vals' finance, legal, tax and Excel benchmarks. Our coding board uses its Code Migration, Vibe Code and ProgramBench results with Terminal-Bench. To see why boards differ, read why AI leaderboards disagree.
Frequently asked questions
Who makes the Vals Index?
Vals AI makes it. The company develops benchmarks with domain experts, publishes model results and evaluates AI products such as legal tools. It runs selected public benchmarks too, and its site lists a range of narrower indexes.
Is the Vals Index a test of AI agents or AI models?
Mostly models doing agent-style work. Vals says it tests at three levels: models through a fixed harness, agents and scaffolds, and finished products. The index ranks models, so read our model vs harness guide before assuming the result transfers to a tool you buy.
Why is finance weighted so heavily?
Because the weights follow GDP share, using government data. Finance and insurance is a much larger share of the US economy than legal or tax services. If that does not match your business, look at the individual sector scores.
Which model is best on the Vals Index?
The list can change with each model release and version, so we do not state it here. See the Vals page for the source, or our leaderboard for the weekly average.
