Vals AI: How Its Benchmarks Work and How to Read Them
Vals AI is a company that benchmarks AI models on work with economic value, such as legal research, financial analysis, tax and coding. Instead of asking whether a model can pass an exam, it asks whether the model can do the job, and it publishes the results on public leaderboards.
The quick answer: Vals is one of the more useful sources for comparing AI models on practical work, especially in law and finance. Treat its scores as one input, though. A high rank doesn’t mean a model can do a job unsupervised, and the right choice still depends on your task, budget and tolerance for error.
Vals AI at a glance
Checked against Vals’ site and press coverage on November 2, 2026.
| Vals AI | |
|---|---|
| What it is | Independent benchmarks and evaluations of AI on real-world tasks |
| Founders | Rayan Krishnan (CEO) and Langston Nashold (CTO) |
| Funding | $40 million Series A led by Andreessen Horowitz |
| Main benchmark | The Vals Index, built from several task benchmarks |
| Business model | Charges companies for evaluations |
| Best for | Buyers and builders choosing models for professional work |
What Vals measures
Vals argues that academic benchmarks stopped being useful. As Andreessen Horowitz puts it, public datasets get saturated, leak into training data, or become targets that models are optimized against.
So Vals works with domain experts to turn real workflows into tests, then grades the finished work against an expert standard. Its public benchmarks include:
- Finance Agent: analyzing complex financial documents.
- Legal Research Bench: finding and applying the law.
- Tax Agent Bench: working through tax questions.
- Terminal-Bench and Vibe Code Bench: coding and building working apps.
- Harvey’s Legal Agent Benchmark: long legal tasks, built by Harvey with Vals running the public leaderboard.
- Immigration Bench: 262 expert-written US immigration questions.
- SAGE: education and grading.
Vals retires tests when models saturate them. In May it swapped CorpFin for an Excel modeling benchmark after CorpFin stopped separating the leaders.
Who leads the Vals Index right now
As of October 30, 2026, showing the best model from each lab:
| Model | Vals Index score |
|---|---|
| Claude Sonnet 5.5 | 70.85% |
| Gemini 4 Carbon | 70.45% |
| Wonder Full | 66.79% |
| GPT-6 Astra | 65.02% |
| Kimi K3.1 Pro | 62.52% |
| DeepSeek V4.1 Flash | 61.26% |
The gaps at the top are small, and new models arrive constantly, so check the live board. Vals also shows cost per test. Gemini 4 Carbon cost $10.51 per test, while GPT-6.1 Terra ranked 10th at $2.92. For many buyers, that trade-off matters more than the top spot.
Results worth a closer look
The individual reports are often more useful than the headline ranking.
Immigration Bench is a good example. Claude Fable 5.1 led with 75.4% average accuracy, but it gave a fully correct answer on only 26.3% of questions. Models met just 7.3% of rubric items on rules from 2025 and 2026, against 64.0% on established law. Vals also found the weakest answers went to the most vulnerable users. That’s a much more useful lesson than a rank: AI is weakest on recent rules, where a mistake can cost the most.
Vals also runs unusual stress tests, like an impossible train refund that showed how models behave when a task can’t be done.
How the legal report built its reputation
Vals made its name in law. Its Legal AI Report (VLAIR), published in February 2025, compared legal AI tools with a group of human lawyers on seven tasks. Harvey scored highest, followed by CoCounsel, and the tools beat the lawyers on document analysis, information retrieval and data extraction.
The design also revealed a limitation. Vendors chose to take part and could withdraw after seeing their results, as long as the withdrawal was made public. LexisNexis withdrew, so its results were left out.
The limits to keep in mind
- Private tests. Vals keeps test sets private to prevent gaming. That’s sensible, but outsiders can’t fully check the questions or grading.
- Who pays. Companies pay Vals to test their models. CEO Rayan Krishnan compares this to a student paying to take the SAT. Andreessen Horowitz compares Vals to credit raters and auditors, fields that also have to manage paid relationships with the firms they assess.
- Benchmark authors. Harvey’s benchmark was built by a major legal AI vendor. LawNext notes this doesn’t make it unreliable, but its definition of good legal work reflects that team’s choices.
- Test settings. Scores depend on settings like reasoning effort. Your own setup may perform differently.
- Rubrics aren’t real life. A task score doesn’t capture how a model works inside your data, tools and review process.
How to use Vals scores
Start with the benchmark closest to your job, not the overall index. Compare cost and speed along with accuracy. Then shortlist two or three models and run them on 20 or so of your own examples. If a mistake would be expensive, keep a human reviewer in the loop whatever the score says.
Vals raised its $40 million Series A in August 2026. It is working on benchmarks for areas like mental health, cybersecurity and biosecurity, so expect its coverage to widen.
FAQ
What is Vals AI?
Vals AI is a company that evaluates AI models on real-world professional tasks and publishes benchmarks, including the Vals Index.
What is the Vals Index?
It’s a composite score built from several Vals task benchmarks across areas like finance, legal, tax and coding.
Which AI model is best on the Vals Index?
On October 30, 2026, Claude Sonnet 5.5 led the board Vals shows by lab at 70.85%, narrowly ahead of Gemini 4 Carbon at 70.45%. Rankings change often.
Is Vals AI independent?
It runs its own evaluations and is separate from the model labs, but it charges companies for evaluations and runs some benchmarks built by others, so read the method behind each result.
What is VLAIR?
VLAIR is the Vals Legal AI Report, a 2025 study that compared legal AI tools with human lawyers.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
