Vals AI Raises $40M Series A Led by a16z to Build Independent AI Model Benchmarks
2026/08/22 12:06:00
Funding Round Signals Rising Demand for Real-World AI Performance Benchmarks
In mid-August 2026, Vals AI announced a $40 million Series A round at a $400 million valuation, led by Andreessen Horowitz with participation from existing investors 8VC, Pear VC, and Bloomberg Beta, plus new backers HRT Ventures and Next Ladder Ventures. The San Francisco-based company positions itself as an independent evaluator of frontier AI models, constructing benchmarks that assess performance on economically meaningful professional work instead of academic-style tests that have largely saturated. Its results already appear in model cards from OpenAI, Anthropic, Google, Meta, and xAI. Revenue has grown eightfold relative to all of 2025, the customer base has doubled, and the team has tripled within six months.
Alongside the funding, Vals released Vals Smith for custom coding benchmarks drawn from any GitHub repository, the RSI Index developed with CoreWeave to measure recursive self-improvement capabilities, ReverseEngBench created with researchers from Columbia, Tufts, UC Berkeley, and UCLA, and an expanded Vals Index 2.0 that weights performance by sector contributions to U.S. GDP. These moves arrive as enterprises face mounting pressure to adopt AI while lacking reliable ways to quantify return on investment, and as governments seek independent measures of frontier capabilities and cyber risk. The core thesis is that independent, continuously updated professional-task benchmarks have become essential infrastructure for the AI economy, closing the growing gap between vendor-reported leaderboard scores and actual ability to complete multi-step work in finance, law, software engineering, and high-stakes risk domains.
Why Traditional AI Leaderboards Have Lost Predictive Power for Enterprise Decisions
Public academic benchmarks once provided a shared language for tracking model progress, but a February 2026 analysis by researchers at ETH Zurich and Stanford examined 60 widely used large-language-model evaluations and found that 29 of them had saturated, with performance gaps between the top and second models falling inside measurement noise. Contamination occurs when test data enters training corpora, sometimes through Common Crawl, while labs also optimize directly against known targets. The result is that high scores on MMLU, HumanEval, or similar suites no longer reliably predict success on messy, multi-step professional workflows. Vals addresses this by keeping primary test sets private, partnering with domain experts to define representative tasks, and retiring benchmarks once they cease to differentiate models.
In May 2026, the company replaced its earlier CorpFin evaluation with a harder Excel Modelling Benchmark precisely because differentiation had collapsed. This adaptive approach treats evaluation itself as continuous infrastructure rather than a static exam, giving enterprises a clearer signal when selecting models for production deployment. Official documentation on the company’s methodology and the a16z investment announcement both emphasize that private, expert-curated sets with ongoing replacement resist the saturation pathway more effectively than either fully public or purely private static tests.
How Vals Constructs Benchmarks Around Actual Professional Workflows
Vals partners with practicing lawyers, financial analysts, software engineers, and clinicians to map the taxonomy of tasks that define competent work in each field, then develops automated grading systems that score generated work product against expert standards rather than multiple-choice correctness or exact string matches. A typical Finance Agent v2 task requires an agent to retrieve supporting data from documents, synthesize relative-value models across peer companies, identify sector catalysts, and produce grounded investment recommendations without hallucinating figures or losing context across steps. In May 2026, the highest-scoring model reached only 52 percent accuracy on this suite, indicating that even frontier systems still fail roughly half of the tasks a professional analyst is expected to handle.
The same philosophy applies to legal research, code migration, terminal-based engineering, and medical coding. Because grading is automated yet calibrated to expert judgment, evaluations can be produced within hours of receiving model access, matching the current weekly cadence of frontier releases. The company has open-sourced portions of its evaluation infrastructure to support reproducibility while protecting the integrity of the private test sets that generate the primary scores. This combination of domain partnership, automated expert-level scoring, and rapid turnaround distinguishes the platform from both academic leaderboards and purely preference-based arenas.
a16z’s Rationale for Leading the Round at a $400 Million Valuation
Andreessen Horowitz framed the investment around the classic information-asymmetry problem described by economist George Akerlof: when sellers know more about product quality than buyers and have incentives to present favorable data, markets degrade toward lower-quality outcomes. Jennifer Li and colleagues at a16z compared the need for an independent AI scorekeeper to the historical emergence of Moody’s in credit markets and independent auditors in public-company reporting. Vals’ revenue model, charging for evaluation services rather than certifying any particular lab’s claims, aims to preserve structural independence.
The firm highlighted the founders’ technical depth and long-standing skepticism about benchmark validity, noting that Rayan Krishnan and Langston Nashold had already collaborated on real-world measurement problems while studying computer science at Stanford. The round’s valuation and the participation of quantitative-trading-linked HRT Ventures further signal that sophisticated capital views reliable measurement as critical infrastructure rather than a peripheral research tool. Official a16z commentary stresses that models are shifting from answering questions to executing multi-hour agentic workflows, raising the cost of a poor model choice from token spend to lost time and customer outcomes.
Founders’ Backgrounds and the Origins of the Independent-Evaluation Thesis
Rayan Krishnan, who serves as the CEO of the company, has a notable background that includes previous work at Palantir, a well-known data analytics firm. His co-founder, Langston Nashold, holds the position of CTO and brings a wealth of experience from his time at major tech companies such as Meta, NVIDIA, and Hudson River Trading. The two founders first crossed paths during a pre-orientation backpacking trip at Stanford University, where they quickly formed a bond. They later collaborated on various computer science coursework, which led them to a significant realization: large language models have the potential to revolutionize the way work is conducted. However, they also recognized that the industry was still lacking credible and effective methods to test these models. Motivated by their insights, they made the bold decision to leave their respective graduate programs in order to establish Vals. The initial focus of their new venture was on finance and coding tasks, which they identified as representing substantial portions of economic activity within the United States.
As they began to gain traction in the market, they received notable citations in major model cards and garnered support for their initiatives from the Department of Commerce, as well as involvement in congressional efforts related to AI policy. The founders place a strong emphasis on the importance of continuously retiring outdated benchmarks and utilizing private test sets. This approach is a direct result of their observations regarding how quickly models can reach the peak of static benchmarks and how training data can influence public evaluations. Their combined experience in both production systems and quantitative environments has significantly informed the design of the evaluation stack, as well as the commercial product that enterprises are now utilizing to select and monitor models in alignment with the cutting edge of technology.
Finance Agent Benchmarks Reveal Persistent Gaps Between Leaderboard Scores and Analyst-Level Work
Even though models achieve near-perfect results on older academic suites, the Finance Agent v2 suite continues to reveal significant material shortfalls that cannot be overlooked. The tasks involved in this suite encompass multi-document synthesis, relative-value modeling, and recommendation generation, all of which closely mirror the daily output produced by professional financial analysts in the industry. The 52 percent ceiling recorded in May 2026 for the leading system serves as a stark reminder that high general-knowledge scores do not automatically equate to reliable and autonomous financial analysis capabilities.
Vals strategically weights its composite Vals Index by considering approximate sector contributions to the U.S. GDP, which gives finance a substantial multiplier effect, ensuring that economic impact is accurately reflected in the aggregate ranking of various models. Enterprises that have successfully deployed models selected partly through Vals evaluations report utilizing the platform not only for the initial selection process but also for the ongoing measurement of product performance against established frontier baselines. Moreover, the existing gap in performance also plays a crucial role in informing capital-allocation decisions within financial institutions. These institutions must justify their AI expenditures with quantifiable productivity gains rather than relying solely on qualitative demonstrations, which can often be subjective and less reliable. This ongoing evaluation process is essential for ensuring that investments in AI technologies yield tangible benefits and improvements in operational efficiency.
Vals Smith Opens Custom Coding Benchmarks Drawn Directly from Enterprise Repositories
With the Series A announcement, Vals made Smith generally available, allowing any organization to generate a custom coding benchmark from its own GitHub repositories. The system extracts real development tasks from historical pull requests and applies hidden tests to determine whether a model can complete the work. Users receive 120 free credits to begin. For companies whose codebases contain proprietary patterns, libraries, and architectural conventions, the distinction between generic Python or Java proficiency and the ability to operate inside that specific environment is decisive.
Smith therefore converts the abstract question of coding capability into a concrete, organization-specific measurement. Early enterprise adopters can now rank models according to performance on their actual engineering workload rather than on public suites that may have been partially memorized or optimized against. The release expands Vals’ reach beyond pre-built domain benchmarks into tailored evaluation infrastructure that scales with customer needs.
RSI Index and ReverseEngBench Expand Coverage into Frontier Risk Domains
In conjunction with the recent funding, Vals has proudly unveiled the RSI Index, developed in collaboration with CoreWeave. This innovative index aims to assess whether various models are capable of executing the research tasks that are essential for advancing model development, compression techniques, training methodologies, parameter optimization strategies, harness engineering practices, and post-training evaluations. Furthermore, the company has introduced ReverseEngBench, a project that was meticulously developed in partnership with esteemed academic institutions, including Columbia University, Tufts University, UC Berkeley, and UCLA. This comprehensive benchmark consists of 19 proprietary programs, which collectively average over 16,000 lines of code.
These programs span a diverse range of areas, including network protocols, firmware, gaming applications, file-format recovery processes, and malware analysis, all of which are safeguarded by an extensive anti-analysis suite designed to enhance security. Preliminary results indicate a significant differentiation among the leading frontier models, revealing considerable remaining potential before agents can consistently reverse-engineer realistic binaries. In addition to these efforts, early-stage work has also been initiated in the realm of mental health applications, with plans for further expansion into critical areas such as environmental impact assessments, military applications, and biosecurity domains. These evaluations, which are oriented towards risk, address capabilities that are crucial for both enhancing enterprise security postures and governmental assessments of cutting-edge systems.
Vals Index 2.0 Aggregates Performance Weighted by Economic Contribution
The redesigned website and Vals Index 2.0 now feature significantly broader coverage across various domains, including finance, coding, and legal tasks. The sector weights utilized in the Index are derived from the value-added data provided by the Bureau of Economic Analysis. The composite formula employed averages the performance metrics within each sector and subsequently combines these sector scores based on their approximate shares of GDP. As of mid-August 2026, Claude Opus 5 is at the forefront of the Index, closely followed by Claude Fable 5 and GPT-5.6 Sol.
The Index is updated on a frequent basis to reflect the latest model releases and serves as a singular headline metric that indicates potential economic impact, while still allowing users to drill down into individual domain leaderboards for more detailed insights. Due to the fact that the underlying benchmarks remain largely private and are refreshed periodically, the Index is able to maintain its differentiation for a longer duration compared to purely public composites. Enterprises and researchers alike consult this Index not only for relative ranking purposes but also for conducting cost-latency-capability trade-off analyses across the current model ecosystem, ensuring they make informed decisions based on the most relevant data available.
Enterprise Adoption Patterns and the Demand for Quantifiable ROI
Large-scale AI deployments are increasingly integrating Vals evaluations when selecting base models and when assessing internal products against frontier performance metrics. This combination of rapid turnaround, domain specificity, and independence effectively addresses the practical challenges that procurement teams and technical leaders encounter: vendor scorecards alone are no longer adequate for making high-stakes decisions.
The remarkable revenue growth of eight times relative to the entirety of 2025, along with a doubled customer base and a tripled headcount within just six months, clearly indicates that demand has transitioned beyond initial experimentation into a phase of operational reliance. Governments have also engaged with the platform to gain valuable insights into capability measurement and cyber risk, thereby reinforcing its significance in both commercial and policy contexts. The platform’s capacity to keep pace with weekly model releases ensures that the signal remains current and relevant, rather than lagging behind the frontier by several months.
Ecosystem and Structural Differentiation from Preference-Based Platforms
Other evaluation efforts are indeed present in the landscape, including preference-voting arenas that aggregate human judgments on open-ended outputs, as well as psychometric approaches that aim to shorten evaluation time significantly. Vals sets itself apart through its expert-curated professional-task design, which ensures high relevance and applicability, alongside automated grading that is meticulously calibrated to meet established domain standards. The company utilizes private test sets that are continuously retired to maintain the integrity of its evaluations, and it boasts a demonstrated citation record by every major frontier laboratory in the field.
This citation record serves as a crucial form of legitimacy that newer entrants must strive to earn in order to gain credibility. Furthermore, the company’s emphasis on economically weighted, multi-step workflows, rather than merely focusing on conversational preference or pure speed, positions it as a vital infrastructure for making informed production decisions, rather than just serving as a player in research leaderboards. Investors have evidently recognized and priced this unique differentiation into the impressive $400 million valuation.
Inference for Model Developers and the Pace of Capability Measurement
Frontier laboratories now operate in an environment where independent scores carry external credibility that self-reported numbers increasingly lack. Citation in model cards signals to enterprise buyers that results have been subjected to third-party scrutiny. At the same time, the continuous creation of harder benchmarks raises the bar that successive models must clear. The “hill-climbing” process that has driven AI progress therefore faces a steeper and more frequently refreshed set of hills.
Developers who treat evaluation as an afterthought risk discovering capability gaps only after deployment; those who integrate independent measurement earlier can prioritize improvements that matter for real work. The open-sourcing of selected infrastructure components further encourages reproducibility while protecting the core private evaluations that generate the highest-signal scores.
Broader Market Need for Trusted Measurement Institutions
As AI systems move deeper into legally and financially consequential workflows, the absence of independent measurement creates the same information-asymmetry risks that historically produced ratings agencies and auditors in other industries. Vals’ growth metrics and the caliber of its backers suggest that capital recognizes this institutional gap. The company’s stated mission, to serve as the independent evaluator of artificial intelligence, aligns with the practical requirements of enterprises seeking ROI clarity and of policymakers seeking capability and risk insight.
Continued expansion into additional professional and risk domains will test whether the methodology scales while preserving independence and signal quality. For now, the combination of recent product releases, quick commercial traction, and high-profile capital commitment establishes Vals as a central reference point for anyone who must decide which models can actually perform the work that matters.
Frequently Asked Questions
What exactly did Vals AI announce in its Series A funding round?
Vals AI closed a $40 million Series A at a $400 million valuation on August 13, 2026. Andreessen Horowitz led the round, with existing investors 8VC, Pear VC, and Bloomberg Beta returning and new investors HRT Ventures and Next Ladder Ventures joining. The company simultaneously released Vals Smith for custom coding benchmarks, the RSI Index, ReverseEngBench, an expanded Vals Index 2.0, and a rebuilt website. Official statements from both Vals and a16z confirm the figures and the accompanying product launches.
How do Vals benchmarks differ from public leaderboards such as MMLU or SWE-bench?
Vals focuses on multi-step professional workflows defined with domain experts and scored by automated systems calibrated to expert standards. Primary test sets remain private and are retired when they stop differentiating models, reducing contamination and optimization gaming. Public academic suites often saturate or leak into training data, limiting their predictive value for real enterprise tasks. Vals’ approach produces results within hours of model access and already appears in major laboratories’ model cards.
What is the significance of the 52 percent score on Finance Agent tasks?
In May 2026, the leading model achieved approximately 52 percent accuracy on the Finance Agent v2 suite, which requires relative-value modeling, document synthesis, and grounded recommendations. The figure indicates that even the strongest available systems still fail roughly half of the tasks expected of a professional financial analyst. This gap between academic scores and professional-task performance is precisely the problem Vals aims to measure and thereby help close.
Who are the founders and what experience do they bring?
CEO Rayan Krishnan previously worked at Palantir. CTO Langston Nashold has experience at Meta, NVIDIA, and Hudson River Trading. Both studied computer science at Stanford and began collaborating on measurement problems before founding the company. Their backgrounds combine production systems experience with quantitative and research perspectives that shape Vals’ emphasis on rigorous, adaptive evaluation.
What does Vals Smith enable for enterprises?
Vals Smith allows any organization to generate a custom coding benchmark directly from its GitHub repositories. The system extracts realistic development tasks from historical pull requests and applies hidden tests. Users start with 120 free credits. The tool converts generic coding benchmarks into organization-specific measurements that reflect proprietary code patterns and engineering practices.
How does the Vals Index incorporate economic weighting?
The Index aggregates performance across finance, coding, and legal tasks, then weights the sector averages according to their approximate contributions to U.S. GDP using Bureau of Economic Analysis data. The resulting composite provides a single headline measure of potential economic impact while still permitting detailed inspection of individual domain results. Version 2.0 expanded coverage and is updated frequently to reflect new model releases.
🔥 KuCoin Offers A More Stable Option in A Volatile Market
If you worry about the frequent ups and downs in the market, and pursue a more stable option to earn money passively, KuCoin is the right place to come:
Simple Earn: Deposit and withdraw tokens anytime, earning stable returns.
Kucoin Earn: Earn stable profits with professional asset management.
Hold to Earn: Earn rewards by holding assets in Funding, Trading, Margin, Futures, Mining, and Unified Accounts.
Staking: Unlock the earning potential of on-chain assets.
Advanced Investments: Advanced Investments offer a variety of structured products to help your money grow in any market.
Shark Fin: Principal Protection and Guaranteed Gains
Dual Investment: Buy low and sell high with transparent return calculations.
Snowball: High yields, with price protection.
Discount Buy: Buy crypto at discount prices.
KCS Loyalty: Level up to enjoy exclusive perks by staking ≥ 1 KCS.
KuCoin Wealth: Discover future value and begin your smart investing journey.
KCS Benefits: Hold and stake KCS to access benefits across the platform.
KCS Staking 2.0: Participate in KCS on-chain governance to earn yield.
Disclaimer: This content is for informational purposes only and does not constitute investment advice. Cryptocurrency investments carry risk. Please do your own research (DYOR).