Vals AI
Independent AI evaluation and benchmarking for large language models on real-world professional tasks.
Website: https://www.vals.ai/home
Cover Block
From the public record
| Field | Value |
|---|---|
| Name | Vals AI |
| Tagline | Independent AI evaluation and benchmarking for large language models on real-world professional tasks. |
| Headquarters | San Francisco, United States |
| Founded | 2024 |
| Stage | Series A |
| Business model | SaaS |
| Industry | Deeptech |
| Technology | AI / Machine Learning |
| Geography | North America |
| Growth profile | Venture Scale |
| Founding team | Co-Founders (2) |
| Funding label | Series A |
| Total disclosed funding | ~$40,000,000 [TechCrunch, September 2026] |
Links
From the public record
- Website: https://www.vals.ai/home
- LinkedIn: https://www.linkedin.com/company/vals-ai
The Short Version
PUBLIC Vals AI builds independent benchmarks for large language models and AI applications on real-world professional tasks, and it is worth investor attention now because it sits in a growing control layer between model vendors and enterprise buyers at a moment when benchmark credibility is becoming commercially and politically salient [TechCrunch, September 2026] [vals.ai] [a16z Policy, August 2026]. Founded in 2024 in San Francisco, the company appears to have moved quickly from an initial seed round in 2025 to a $40 million Series A announced in August 2026, led by Andreessen Horowitz, with earlier backing from 8VC and Bloomberg Beta [TechCrunch, September 2026] [Crunchbase] [Dealroom.co].
The product pitch is straightforward: Vals tests models on work-like tasks across domains such as finance, software, law, and medicine, and its public materials emphasize private datasets, in-house evaluations, model comparisons, and an enterprise product called Vals Smith for customer-specific evaluation and model selection [vals.ai] [TechCrunch, September 2026]. That positioning matters because the company is not trying to win at model training itself; it is trying to become a trusted measurement layer, which can be valuable if enterprises increasingly need third-party evidence on accuracy, latency, cost, and benchmark leakage rather than vendor claims alone [The New York Times, May 2026] [a16z, September 2026].
The founding bench is young but relevant to the product category. CEO Rayan Krishnan studied computer science at Stanford, previously interned at Palantir, worked at Microsoft and Stanford's AI lab as an undergraduate, and has been described in public coverage as having a background in benchmarks and evaluations; co-founder Langston Nashold also studied computer science at Stanford [TechCrunch, September 2026] [Teahose, September 2026]. Public reporting also points to fast hiring, with headcount growing from eight employees to roughly 30 during 2026 and third-party directories showing the company in the mid-30s by late September, which is consistent with an early company scaling around a fresh round, though the exact count varies by source [CNBC, October 2026] [TipRanks.com].
The financing picture is comparatively clear for a company at this age. Public sources show an undisclosed seed in 2025 and a $40 million Series A in August 2026, while multiple sources also associate the round with a reported $400 million valuation; the business model is SaaS, but public evidence on customer mix, contract size, and renewal behavior remains thin [TechCrunch, September 2026] [Wilson Sonsini] [Crunchbase]. Over the next 12 to 18 months, the central question is whether Vals can turn benchmark visibility into durable enterprise workflow adoption, particularly through Vals Smith and domain-specific evaluation products, without losing its claim to independence as foundation model labs and large enterprises become more demanding buyers of evaluation infrastructure [vals.ai] [a16z Policy, August 2026] [The New York Times, December 2025].
Single-source, plausible -- Section relies on a mix of independent reporting, company materials, and directory sources; funding is well corroborated, while product scope and operating metrics are only partially corroborated.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Series A |
| Business Model | SaaS |
| Industry / Vertical | Deeptech |
| Technology Type | AI / Machine Learning |
| Geography | North America |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (2) |
| Funding | Total disclosed funding of approximately $40,000,000 |
The Company in Brief
PUBLIC
Vals AI appears to be moving quickly into a narrow but increasingly consequential layer of the AI stack: independent evaluation. Public company materials describe the business in plain terms as benchmarking AI models on real-world work, and Crunchbase lists the company as founded in 2024 and based in San Francisco [vals.ai, retrieved 2026] [Crunchbase, retrieved 2026]. The reported operating premise is straightforward: rather than rely on broad public benchmarks, Vals tests models on task-specific workflows and publishes comparative results through its benchmark and model-comparison surfaces [vals.ai, retrieved 2026].
The public chronology is short but already useful. Crunchbase identifies Vals AI as a 2024-founded company, while the company website shows a product footprint that includes Benchmarks, Models, Comparison, Vals Smith, Reports, and Government navigation, suggesting a business that has expanded beyond a single leaderboard into a broader evaluation platform by 2026 [Crunchbase, retrieved 2026] [vals.ai, retrieved 2026]. A Dealroom item tied to the company records a $40 million Series A in August 2026, which aligns with later database coverage that the company had reached the Series A stage [Dealroom.co, retrieved 2026] [Crunchbase, retrieved 2026].
What remains less clear from public corporate records is the legal-entity detail. The available Crunchbase profile and company website establish the founding year, headquarters, and product framing, but they do not surface a legal entity name or incorporation filing in the materials reviewed here [Crunchbase, retrieved 2026] [vals.ai, retrieved 2026]. For now, the cleanest read is that Vals is an early but well-capitalized San Francisco startup building evaluation infrastructure for model buyers, model builders, and adjacent institutional users, with most of the public evidence concentrated in company-controlled materials and venture databases [vals.ai, retrieved 2026] [Crunchbase, retrieved 2026] [Dealroom.co, retrieved 2026].
Single-source, plausible -- Grounded primarily in Crunchbase, Dealroom, and the company website; headquarters, founding year, stage, and product framing are corroborated, but legal entity details were not confirmed from public filings used in this section.
What They Have Built
Mixed sourcing
Vals AI is selling a narrower product than the broad "AI infrastructure" label suggests. The public record points to an evaluation layer for large language models and AI applications, with the company positioning itself around testing models on real-world work rather than on academic or public benchmark sets [vals.ai, retrieved 2026] [TechCrunch, September 2026] [The New York Times, May 2026]. On its website, Vals says it independently tests models on economically valuable tasks across finance, software, law, and frontier risk domains, and publishes comparison views that include performance, cost, latency, and token usage [vals.ai, retrieved 2026].
The evidence for product breadth is strongest on model benchmarking and weaker on enterprise workflow claims. Vals' homepage shows benchmark, model comparison, report, and Vals Smith product surfaces, and says the company creates many benchmarks in-house and runs its own evaluations [vals.ai, retrieved 2026]. Separate company materials say Vals runs 32 independent benchmarks and offers domain-specific benchmarking in areas such as legal, tax, and finance, including a legal reasoning leaderboard; those details are company-sourced and should be read as product claims rather than independently verified adoption signals [vals.ai, retrieved 2026]. Press coverage adds one concrete enterprise use case: TechTimes reported that Vals Smith lets customers build custom coding benchmarks from their own GitHub repositories [TechTimes, August 2026].
The technical differentiation, as publicly described, rests less on a novel model layer and more on benchmark design. Company materials and secondary coverage describe private datasets, held-out tests, and domain-specific leaderboards intended to reduce test leakage, while evaluating dimensions such as reasoning, tool use, multimodality, accuracy, latency, and cost [vals.ai, retrieved 2026] [TechCrunch, September 2026]. That framing is directionally supported by outside references to Vals tracking or testing the latest AI systems, and by examples of published benchmark results on named models, but the underlying methodology, scoring consistency across domains, and repeatability of customer-specific evaluations are not fully documented in the public record [The New York Times, December 2025] [The New York Times, May 2026] [BigGo Finance, retrieved 2026].
Single-source, plausible -- Core product positioning is corroborated by the company website and TechCrunch, but many feature-level details remain company-sourced or only partially verified by secondary coverage.
Market Size and Demand
PUBLIC
The market matters now because enterprises are moving from experimenting with foundation models to selecting, governing, and re-testing them in production, which raises the value of independent evaluation from a research accessory to an operating requirement [The New York Times, December 2025] [The New York Times, May 2026] [a16z Policy, August 2026].
Public evidence here is thinner on formal market sizing than on demand formation, so the cleaner read is qualitative rather than numeric. Named-publisher coverage places Vals inside the fast-expanding layer that measures model performance on real work, not just on public benchmark exams [TechCrunch, September 2026] [The New York Times, December 2025]. That distinction matters because model quality has become harder for buyers to compare as vendors ship frequent updates, pricing changes, and modality expansions, while enterprises still need use-case-specific evidence before deployment [The New York Times, July 2026] [a16z, September 2026].
Demand drivers are visible in the way mainstream coverage now discusses model competition. The New York Times described Vals as tracking or testing the latest AI technologies in reporting across 2025 and 2026, which suggests the company is operating in a market where benchmark visibility has become part of the model procurement and media cycle, not just internal lab practice [The New York Times, December 2025] [The New York Times, May 2026]. TechCrunch's September 2026 reporting adds the commercial angle: Vals is positioning itself around evaluations on professional tasks and customer-specific workflows, a wedge that becomes more relevant as companies choose among increasingly capable models for coding, legal, financial, and healthcare use cases [TechCrunch, September 2026].
The closest adjacent markets are broader AI observability, model governance, safety testing, and enterprise software procurement tools. Vals' own public materials frame the product around benchmark results, model comparisons, cost, latency, and domain-specific testing, which places it near both benchmarking infrastructure and decision-support software for AI buyers [vals.ai] [a16z Policy, August 2026]. Substitute behavior also exists: large model labs can run internal evals, and enterprises can build bespoke testing suites themselves, so the addressable segment is likely strongest where customers want a third party, shared methodology, or faster time to deployment decisions [a16z, September 2026] [TechCrunch, September 2026].
Regulatory and macro forces also support the category, even if the exact spend pool is not quantified in the cited sources. The policy discussion hosted by a16z in August 2026 centers on what benchmarks can teach policymakers and how government may use evaluation tools, which points to a compliance and public-sector tailwind beyond pure enterprise purchasing [a16z Policy, August 2026]. Bloomberg's September 2026 reporting on AI's environmental impact per task adds a second macro pressure: as model complexity rises, buyers have stronger reason to compare systems not only on capability but also on efficiency and operating cost, areas that benchmarking products can make legible [Bloomberg, September 2026].
| Market lens | Public evidence | Read-through for Vals |
|---|---|---|
| Enterprise model selection | Buyers are comparing rapidly changing frontier models for production tasks [The New York Times, July 2026] [TechCrunch, September 2026] | Supports demand for third-party evaluations tied to workflows rather than static test sets |
| Governance and policy | Benchmarking is being discussed in policy and government-use contexts [a16z Policy, August 2026] | Expands the category beyond private R&D and into oversight, procurement, and standards |
| Cost and efficiency measurement | AI task complexity is raising environmental and compute cost scrutiny [Bloomberg, September 2026] | Makes latency, cost, and token-use comparisons more commercially relevant |
| Domain-specific deployment | Public materials emphasize finance, law, coding, and other professional tasks [vals.ai] [TechCrunch, September 2026] | Suggests the near-term market is segmented by regulated or high-value workflows rather than generic chatbot usage |
The table points to a category that is broad in need but still early in budget definition. For Vals, that usually means the market case rests less on a published TAM figure and more on whether independent evaluation becomes embedded in enterprise AI purchasing and governance motions.
Single-source, plausible -- Section rests on named-publisher coverage from The New York Times, TechCrunch, Bloomberg, and a16z Policy, plus company website product framing. No named third-party market size report was identified in the provided public sources.
Who Else Is Fighting for This
MIXED Vals AI is positioning itself less as a model builder and more as an independent measurement layer, which places it adjacent to frontier labs and enterprise AI tooling rather than in direct feature competition with either group [TechCrunch, September 2026] [vals.ai].
The public record is thin on named peers, so the competitive map is clearer by function than by logo. On one side sit the foundation model companies whose products Vals measures, including Meta and the broader set of labs represented on its public benchmark pages, such as Anthropic, Google, OpenAI, xAI, Mistral, DeepSeek, Qwen, MiniMax, and others shown in Vals' comparison interface [vals.ai] [The New York Times, May 2026]. Those firms are not direct substitutes for Vals' service, but they do control the underlying model supply and publish their own evaluations, which makes them partial competitors for mindshare around what counts as model quality [a16z, September 2026] [a16z Policy, August 2026]. On the other side are enterprise buyers deciding which model to use for legal, finance, coding, or other professional workflows, a decision set that can be served by an outside evaluator like Vals or handled internally by the buyer's own team [TechCrunch, September 2026] [vals.ai].
Vals' current edge appears to rest on independence claims, proprietary test design, and a product surface that turns evaluation into an operating tool rather than a one-off research artifact. The company says it runs 32 independent benchmarks, creates many benchmarks in-house, and uses private or held-out task sets to reduce test leakage, while its Vals Smith product is positioned to help enterprises choose the right model for a specific workflow [vals.ai] [TechCrunch, September 2026]. That combination matters because benchmark credibility tends to erode once test sets become widely known, but it is also a perishable edge: if large labs improve their own domain evaluations or enterprises decide their internal data is the only benchmark that matters, the independent layer can be compressed unless Vals keeps refreshing datasets and building workflow-level trust faster than model vendors can replicate it [a16z, September 2026] [The New York Times, December 2025].
The exposure is straightforward. Andreessen Horowitz is both a source of capital and a signal amplifier, but distribution still appears concentrated in founder-led evangelism and public benchmark visibility rather than an owned channel with entrenched customer switching costs [TechCrunch, September 2026] [a16z, September 2026]. Meta is a useful example of the pressure point: when Vals' held-out benchmarks suggested Llama 4 underperformed, that result helped Vals' credibility as an independent grader, but it also illustrates how quickly the company can be pulled into a contested narrative where the model developer has larger brand reach, larger research teams, and stronger ability to publish competing evidence [BigGo Finance] [vals.ai]. The same risk applies more broadly to Anthropic, OpenAI, and Google, each of which has the resources to develop proprietary eval stacks or customer-specific testing programs that narrow the need for a third party [The New York Times, May 2026] [The New York Times, July 2026].
Over the next 18 months, the most plausible scenario is not one winner taking the category, but a split market between public benchmark influence and private deployment testing. Vals is the likely winner if enterprise buyers and policymakers continue to prefer an outside party to test models on economically meaningful work, especially where auditability and domain-specific comparisons matter more than raw model branding [a16z Policy, August 2026] [TechCrunch, September 2026]. The likely loser if that preference weakens is any independent evaluator whose differentiation is mainly leaderboard visibility rather than embedded workflow software; by contrast, a lab such as Anthropic would be advantaged if buyers decide the best benchmark is direct testing inside the vendor ecosystem, because it already has product distribution and enterprise relevance at the model layer [The New York Times, May 2026] [a16z, September 2026].
Single-source, plausible -- Core positioning and funding are corroborated by TechCrunch and the company website, but named direct competitors are not established in the source set and several competitive inferences rely on company materials and category analysis.
Opportunity
PUBLIC The prize here is unusually large because a trusted evaluation layer for AI could sit between every major model vendor and every enterprise buyer, which would place Vals AI in the path of both model selection and ongoing governance if adoption broadens from point benchmarks to repeatable operating infrastructure [TechCrunch, September 2026] [vals.ai].
The headline opportunity is straightforward: Vals could become the default independent testing and benchmarking layer for high-stakes AI deployments. That is a bigger outcome than selling one-off model scorecards, and the public evidence makes it at least reachable. The company has raised a $40 million Series A led by Andreessen Horowitz after an earlier seed led by 8VC and Bloomberg Beta, which suggests investors see a category with room for a scaled platform rather than a narrow tooling product [TechCrunch, September 2026]. The product positioning is also pointed at a real market failure. Public benchmarks are easy for model vendors to optimize against, while buyers in law, finance, coding, and healthcare need evidence on actual tasks, latency, and cost, which is close to how Vals describes its work and how outside coverage has framed the company [vals.ai] [The New York Times, December 2025] [The New York Times, May 2026].
A second reason the upside is not purely aspirational is that the company appears to be widening from public benchmark publishing into enterprise software. Vals markets Vals Smith as an application for model selection, and outside coverage says the product can help customers build custom coding benchmarks from their own GitHub repositories, which moves the company closer to workflow integration and away from media-style rankings alone [vals.ai] [TechTimes, August 2026]. If that motion works, Vals is not only grading models. It becomes part of how organizations decide what to deploy, monitor changes over time, and justify those decisions internally.
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Evaluation system of record | Vals becomes the standard platform enterprises use to test and compare models before deployment and after major model updates | Vals Smith gains adoption as teams need repeatable, domain-specific model selection workflows [vals.ai] [TechTimes, August 2026] | The company already presents an enterprise product for model choice, and its public site emphasizes comparisons across performance, cost, latency, and token usage rather than a single benchmark score [vals.ai] |
| Independent layer for frontier model launches | Major model labs and buyers increasingly reference Vals benchmarks when new models ship, making Vals a default third-party credibility layer | Continued model release velocity and visible benchmark miss cases such as reported underperformance by major models on held-out tests [The New York Times, July 2026] [BigGo Finance] | Vals is already publishing held-out benchmark results and is cited in mainstream reporting about AI model performance, which indicates early relevance in the model-release cycle [The New York Times, May 2026] [The New York Times, December 2025] |
| Policy and regulated-sector standard | Government and regulated industries adopt third-party benchmark frameworks as part of procurement, oversight, or deployment approval | Benchmarking becomes part of AI governance discussions, helped by Vals' visibility in policy conversations [a16z Policy, August 2026] | The company has already appeared in a policy-focused discussion on AI benchmarks and government use, while its site explicitly covers government and high-stakes domains such as law, finance, and frontier risk [a16z Policy, August 2026] [vals.ai] |
The compounding mechanism would come from proprietary evaluation data, workflow embedding, and brand trust reinforcing each other. Each customer-specific benchmark adds more knowledge about how models behave on economically valuable tasks, and the company says it creates many benchmarks in-house while also using private datasets and domain-specific leaderboards to reduce leakage [vals.ai]. Each public model launch then gives Vals another chance to publish comparative evidence, attract enterprise traffic, and convert that interest into custom evaluation work through Vals Smith [vals.ai] [TechCrunch, September 2026]. If customers begin to store their internal test suites and deployment decisions in the platform, switching costs rise because the historical comparison set becomes part of the product's value, not just the latest score.
The size of the win is large enough to matter even after applying a conservative filter. Public sources report Vals raised its Series A at a $400 million valuation in August 2026, which is already a meaningful entry price for a company founded in 2024 and signals investor belief that evaluation can become a major control point in the AI stack [Wilson Sonsini] [TechCrunch, September 2026]. If the "evaluation system of record" scenario plays out, one credible framing is not a benchmarking publisher but an infrastructure software company serving both model vendors and enterprise buyers. In that case, a multi-billion-dollar outcome is plausible, and a $3 billion to $8 billion range would be a reasonable way to frame the upside (scenario, not a forecast), assuming Vals converts benchmark authority into durable software usage rather than remaining primarily a research surface. That range is an analytical scenario anchored by the company's current $400 million reported valuation and the scale investors are already underwriting, not by disclosed revenue or a public comparable specific to Vals' category [Wilson Sonsini] [TechCrunch, September 2026].
Single-source, plausible -- Section relies on TechCrunch, The New York Times, a16z Policy, Wilson Sonsini, TechTimes, and Vals' website, but several core product-scope claims remain company-described rather than independently customer-verified.
Sources
From the public record
[TechCrunch, September 2026] Vals, backed by Andreessen Horowitz, is looking to become the gold standard for AI benchmarking | https://techcrunch.com/2026/09/19/vals-backed-by-andreessen-horowitz-is-looking-to-become-the-gold-standard-for-ai-benchmarking/
[vals.ai] AI Benchmarks for Real-World Tasks | Vals AI | https://www.vals.ai/home
[a16z Policy, August 2026] Measuring AI: What Policymakers Can Learn from Benchmarks | https://a16zpolicy.substack.com/p/measuring-ai-what-policymakers-can
[Crunchbase] Vals AI - Crunchbase Company Profile & Funding | https://www.crunchbase.com/organization/vals-ai
[Dealroom.co] vals.ai raises $40M Series A at $400M valuation, led by Andreessen Horowitz | https://app.dealroom.co/news/note/vals-ai-raises-40m-series-a-led-by-andreessen-horowitz
[The New York Times, May 2026] Anthropic Tops OpenAI to Become the World’s Most Valuable A.I. Start-Up | https://nytimes.com/2026/05/28/technology/anthropic-claude-opus-48.html
[a16z, September 2026] Who Grades the AI Models? | Ben Horowitz & Rayan Krishnan | Andreessen Horowitz | https://a16z.com/podcast/who-grades-the-ai-models-ben-horowitz-rayan-krishnan/
[Teahose, September 2026] Rayan Krishnan interview or profile | https://www.teahose.com/
[CNBC, October 2026] CNBC coverage cited for Vals AI headcount growth | https://www.cnbc.com/
[TipRanks.com] Vals AI company profile | https://www.tipranks.com/
[Wilson Sonsini] Vals AI job listing or company materials referencing valuation | https://www.wsgr.com/
[Dealroom.co] Vals AI company profile | https://dealroom.co/companies/vals-ai/
[BigGo Finance] BigGo Finance coverage referencing Meta's Llama 4 on Vals benchmarks | https://biggo.com/
[The New York Times, December 2025] Can OpenAI Respond After Google Closes the A.I. Technology Gap? | https://www.nytimes.com/2025/12/11/technology/openai-google-ai-technology-gap.html
[The New York Times, July 2026] OpenAI Releases GPT-5.6 Sol, Its Most Powerful AI Model Yet | https://www.nytimes.com/2026/07/09/technology/openai-sol-ai.html?eafs_enabled=false
[Bloomberg, September 2026] AI’s Environmental Impact Per Task Balloons With More Complexity | https://www.bloomberg.com/news/articles/2026-09-03/ai-s-environmental-impact-per-task-balloons-with-more-complexity?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc4ODQ1Mzk2MywiZXhwIjoxNzg5MDU4NzYzLCJhcnRpY2xlSWQiOiJUS1NCV1pSMjRVOFMwMCIsImJjb25uZWN0SWQiOiJGRjMyOTZDMzVFNEI0QTRBQjFFNUNDMzJDQzMzMUNDQiJ9.8ShIwgSwrkJWvsN5gaj4nXw_Jy9W3gDmfbKFK1Tb6kk
Articles about Vals AI
- Vals AI Tests AI Models on 32 Private Benchmarks for Law and Finance — The Andreessen Horowitz-backed startup sells independent testing to model labs and enterprises, aiming to replace public benchmarks with private, real-world tasks.