Vals AI Raises $40M to Fix Broken AI Benchmarks
Enterprise AI procurement has a benchmarking crisis. Every major AI vendor publishes scores proving their model leads the field. Every score is, on some level, marketing. The startup Vals AI, which raised a $40 million Series A led by Andreessen Horowitz on September 19, 2026, is building the independent evaluation infrastructure that would let enterprise buyers trust a number again.
The Problem: Benchmarks That Reward Gaming
The benchmark contamination problem is not theoretical. When a test is public, model developers can train directly against it, producing systems that score well on the exam without gaining equivalent capability on real professional tasks. Academic benchmarks, many of them designed years before today’s frontier models existed, have become as much a PR tool as a technical signal.
“We were seeing a bunch of new, very capable models come to market quickly, and the academic benchmarks were not keeping up with that frontier advance,” said Rayan Krishnan, Vals’ 25-year-old co-founder, in a September 2026 interview with TechCrunch. Krishnan, who previously interned at Palantir and worked at Microsoft and Stanford’s AI Lab, founded Vals in 2024 specifically to close that gap.
The result is a credibility vacuum for enterprise buyers. A procurement team evaluating a seven-figure inference contract cannot verify whether a vendor’s published scores reflect what the model will do in production, what it will do on the specific legal research workflow or financial analysis task the team actually needs, or how it performs when it fails. Most benchmarks only measure success.
Vals’ Approach: Private Tests, Real Tasks, Failure Penalties
Vals differentiates on two dimensions. First, it keeps its test materials confidential. No one outside the company sees the specific questions, which prevents vendors from training against the exam. Second, rather than measuring abstract intelligence (the “bar exam type test” model that Krishnan explicitly rejects), Vals grades models on whether they can produce work comparable to a human expert in specific professional domains.
The current domain coverage spans law, finance, coding, cybersecurity, biosecurity, mental health, recursive self-improvement, and law-of-armed-conflict scenarios including the Geneva Convention. Vals evaluates both capability and failure modes: not just “can the model do this” but “what happens when it goes wrong in production.”
The Vals Index, the company’s public-facing leaderboard, translates this into a single composite score per model. As of the August 2026 update, Claude Fable 5 scored 75.14% on the index. Enterprise procurement teams are beginning to cite these scores in RFPs as a neutral reference point alongside vendor-supplied benchmarks.
A more detailed breakdown of where the major approaches differ:
| Dimension | Public Academic Benchmark | Vendor Self-Reported | Vals (Private Task-Specific) |
|---|---|---|---|
| Test materials | Public | Proprietary | Proprietary (confidential) |
| Gaming risk | High | Very high | Low |
| Domain specificity | General | Varies | Law, finance, coding, cyber, bio |
| Failure mode coverage | Rare | Rare | Yes |
| Third-party independence | Varies | None | Yes |
| Federal/compliance use | Rarely cited | Not accepted | Federal agency program launched |
The Business Model and the Growth Numbers
Vals charges AI companies to test their own models. The structure is counterintuitive but precedented: Krishnan compares it to students paying the College Board for the SAT. The implicit logic is that a credible test result is worth paying for even if the result is unflattering, because it identifies where to improve and because buyers increasingly require it.
The model appears to be working. Revenue grew eightfold year-over-year in 2026. The company started 2026 with eight employees and had reached 25 by September, with plans to hire 10 to 15 more. Krishnan has also launched a program offering model evaluations to US federal agencies, a customer segment with longer sales cycles but sticky, high-value contracts once secured.
The September 2026 Series A, led by Andreessen Horowitz with participation from earlier backers 8VC and Bloomberg Beta, extends the runway for domain expansion, expert hiring, and the build-out of enterprise continuous monitoring products. The implied bet from a16z: that independent evaluation becomes structural infrastructure for the AI economy, a position analogous to what credit rating agencies occupy in finance or what Gartner has historically occupied in enterprise software.
Why This Matters Now: AI Is Going Public
Krishnan himself articulates the timing with clarity. “AI companies are starting to go public. SpaceX went public. Anthropic is slated for later this year. I suspect OpenAI will be public soon,” he told TechCrunch. “I think as AI models become a core part of the economy and are diffused more broadly, the types of benchmarks and evaluations that we do are going to drive their usage and be a central part of how these companies submit public filings.”
That is the structural shift that makes this round more than a startup funding story. As long as AI was primarily a cost center, procurement teams could rely on gut feel, demos, and vendor reference calls. When AI becomes a material line item in public company financials, the standards for evidence change. An independent, audit-grade evaluation becomes the kind of thing a CFO, a board, or a regulator can ask for and expect to receive.
For enterprise teams already deploying AI agents across GTM, operations, or legal workflows, this shift is already underway. The enterprise AI adoption gap documented in mid-2026 was partly a readiness problem and partly a trust problem. Teams were not sure what models could actually do in their specific context. Vals is, in effect, selling the evidence layer that closes the trust gap.
The Risks Worth Naming
The business model carries a structural tension that The Beat and FourWeekMBA have both flagged in their analyses: Vals is paid by the same companies whose models it evaluates. That relationship is not unusual (credit rating agencies operate the same way) but it invites scrutiny as the stakes increase. If Vals scores begin feeding into public filings or investment decisions, the independence of the examiner becomes a material question.
There is also a model-cycle risk. When the next generation of frontier models ships, leaderboard positions flip and the index’s perceived authority can reset. Vals’ defensive move is to offer enterprises evaluations against their own proprietary data and workflows, creating a more durable signal than any generic benchmark can. Whether a 25-person team can execute that product expansion at the pace the funding multiple implies remains to be seen.
Competition is latent rather than absent. Large AI vendors and larger incumbents could build internal evaluation frameworks or fund open alternatives that undercut Vals’ independence pitch. Regulatory bodies could mandate their own testing regimes.
What Enterprise Teams Should Do Now
The practical implication for enterprise AI buyers is not to wait for Vals to become the universal standard before acting, but to begin building evaluation discipline into procurement now.
Three concrete steps:
- Define task-specific success criteria before evaluating models. The question is not “which model scores highest on MMLU” but “which model produces correct output on the thirty most common tasks your team runs.”
- Request vendor-reported scores alongside independent evaluations. Vals is one option; Salesforce’s Agentic Enterprise Index and similar third-party assessments add useful signal alongside it.
- Test for failure modes, not just capability. Vals’ coverage of negative outcomes reflects what production AI deployment actually looks like. Build your own failure-case library for the models you are evaluating.
If you want to build evaluation infrastructure into your AI operating model before your competitors do, Enera can help.
Sources: TechCrunch, September 19, 2026; The Beat, September 20, 2026; Gokhshtein Media, September 20, 2026; FourWeekMBA, September 19, 2026.