Valz raises $40 million led by Andreessen Horowitz as closed door auditing displaces academic AI tests

Excelling in benchmark model tests is no longer just a technical indicator; it has become a public relations tool used by AI labs to demonstrate the merit of their systems to investors and clients whenever the numbers tip in their favour. However, companies' ability to game those outdated metrics, which were not designed to measure the capabilities of modern models and with systems pre-trained on generic test questions, has exposed the shortcomings of the prevailing system. At this moment, the startup Valz secured a first round funding of forty million dollars led by Andreessen Horowitz, after an earlier seed round led by ETF and Bloomberg Beta, in an effort to establish a new independent standard for evaluating generative model efficiency.

In 2024, Valz was launched by Ryan Krishnan, 25, who had interned at Palantir and conducted research in the AI lab at Stanford University and Microsoft. Krishnan observed that academic tests failed to keep pace with the rapid breakthroughs of advanced frontier models and continued to focus on measuring general theoretical knowledge through abstract methods similar to bar exam assessments. Valz takes a different approach, completely concealing test questions and data from public view to prevent gaming and model training on them, and instead focuses on measuring models' ability to perform complex tasks that emulate the requirements of critical sectors such as law, finance, and programming, to verify that their outputs reach human level work quality and to examine their negative ramifications if released without strict controls.

The suite of tests that the company develops from its San Francisco headquarters is expanding to include unconventional areas such as iterative self improvement of models, cybersecurity, biosafety, the application of international humanitarian law and the Geneva Conventions in armed conflicts, as well as a dedicated program to provide targeted assessments for federal agencies. The company charges developers fees for evaluating their models, a business model that its founder likens to the SAT test fees paid by students. This evaluation allows companies to discover and address faults before commercial launch, driving Valz's revenue to increase eightfold compared with the previous year, and growing its staff from eight individuals at the start of the year to 25 employees, with plans to move to a larger location and add between ten and fifteen new personnel.

This evaluation represents a mandatory pillar expected to regulate corporate disclosures as a wave of public offerings approaches in financial markets.With major tech companies such as SpaceX being listed, ongoing arrangements for the launch of Anthropic, and expectations for the launch of OpenAI, Krishnan believes that independent evaluation standards will become a decisive factor that determines the scale of model usage and supports the official data companies provide to investors to justify their future investments and the reliability of their tools.

This shift in the global evaluation architecture requires technology leaders and corporate procurement officials in the Gulf, Egypt, and the Levant to reset their standards immediately: reliance on public leaderboards and advertised test results in model marketing is no longer a valid guarantee for building software or contractual decisions. Financial institutions and government bodies in the region, which heavily invest in automating legal contracts and credit analyses, face a reality that mandates banning direct publication based on developers' claims and building blind tests tailored to local work data. The critical skill for your engineering team is no longer limited to API integration, but to the ability to audit models through closed environments that accurately measure error probabilities and operational risks before any model is approved for sensitive production lines.