Vals AI Raises $40M From a16z to Benchmark AI Agents on Real Professional Tasks

Vals AI has closed a $40 million Series A at a $400 million valuation, led by Andreessen Horowitz, to build out the infrastructure for evaluating AI models on real professional work — a market that crystallised around one damning finding: the best frontier AI systems currently fail 52% of the tasks a professional finance analyst would be expected to complete.

Why Benchmark Scores No Longer Mean What They Used To

For most of AI's modern history, academic benchmarks gave the industry a common language. Researchers published scores on MMLU for general knowledge, HumanEval for code, GSM8K for math — and those numbers propagated into model cards, press releases and procurement decisions. The system worked until model training began explicitly optimising for benchmark performance rather than the underlying capability the benchmarks were designed to measure. Once that inflection point passed, benchmark scores and real-world usefulness started diverging.

The problem is structural, not incidental. A 2026 preprint from researchers at ETH Zurich and Stanford found that making test sets private does not systematically prevent saturation — once a benchmark's distributional characteristics become widely known through training data and competitive pressure, scores compress regardless of whether the specific questions were ever published. What does resist saturation, the preprint found: expert-curated benchmarks with large test sets that are continuously retired and replaced as models improve. That last finding is a precise description of what Vals has built.

How Vals Actually Grades AI Models

Vals AI works with domain experts — financial analysts, lawyers, software engineers, healthcare professionals — to construct evaluations around multi-step professional workflows. Rather than asking a model to select the right answer from a multiple-choice set, Vals tests whether it can complete the full analytical chain: retrieve the right supporting data, synthesise it without hallucinating figures, and chain reasoning across multiple documents without losing context. Automated grading then scores model outputs against expert-level professional standards.

The company treats its benchmarks as perishable by design. When a benchmark stops differentiating strong models from weaker ones — when saturation sets in — Vals retires it and builds a harder one. In May 2026, it replaced its CorpFin benchmark with a new Excel test after CorpFin stopped providing sufficient separation between frontier models. That lifecycle management — not just the privacy of individual test questions — is the structural fix to benchmark drift. Vals can produce results within hours of receiving access to a new model, which matters because the frontier model release cycle is now measured in weeks rather than years. The company's evaluations have been cited in official model cards from OpenAI, Anthropic, Google, Meta and xAI — effectively achieving the status of a third-party financial auditor for AI capability claims.

Fifty-Two Percent — The Finance Analyst Ceiling

The headline finding from Vals' Finance Agent v2 benchmark — that the best available AI model in May 2026 correctly completed only 52% of tasks a professional analyst would be expected to handle — is the kind of number that cuts through the hype. The tasks tested are not edge cases: they include building relative-value models across peer companies, synthesising sector catalysts, generating investment recommendations grounded in actual document data, and researching SEC filings. A model that scores 52% on those tasks is not a reliable autonomous tool for professional financial work, regardless of its MMLU score.

The Vals Index — which measures agentic model performance across finance, coding and legal tasks, weighted by each sector's share of US GDP — currently ranks Claude Opus 5 first, Claude Fable 5 second and GPT 5.6 Sol third across 46 tested models. A companion RSI Index, launched August 13, measures how close frontier AI is to recursive self-improvement: models run open-ended AI R&D tasks autonomously, scored against human research records. The connection between Anthropic's Fable 5 safeguard architecture and its top-two Vals ranking is not coincidental — safety-tuned models tend to be more reliable on multi-step agentic tasks precisely because they refuse to hallucinate when uncertain.

The Lemon Problem and Why a16z Bet $40M on the Solution

Economists use the term "lemon problem" — from George Akerlof's 1970 paper — to describe markets where buyers cannot distinguish good products from bad ones, causing quality to collapse toward the lowest common denominator. AI has developed a version of this problem: enterprise buyers cannot reliably evaluate whether a model that claims to be a capable coding assistant, legal analyst or financial researcher will actually perform in their environment. Vals is betting that solving this verification layer is structurally valuable enough to support a standalone business at scale.

The round's metrics suggest that bet is already paying off: 8x revenue growth over 2025, customer count doubled, team tripled — all in six months. Returning investors 8VC, Pear VC and Bloomberg Beta were joined by HRT Ventures and Next Ladder Ventures, alongside a16z. The company was founded by Langston Nashold and Rayan Krishna, and operates from San Francisco with approximately 15 employees — a headcount that will expand substantially on the new capital. The broader agentic AI deployment wave that xAI's Grok Bot and the agent plugin ecosystem are building creates exactly the demand environment Vals needs: enterprises increasingly deploying AI agents on real professional workflows, needing independent evidence that those agents will actually perform. As Microsoft's own agentic AI guidance acknowledges, evaluation is the unsolved layer in enterprise AI deployment — and Vals is positioning itself as the auditor that makes deployment decisions defensible.