Mercor raises $750M to test AI agents in simulation

Mercor has completed a Series D funding round of $750 million, bringing its private valuation to $20 billion. The company aims to industrialize AI agent testing on simulated software, supported by a vast network of evaluators. However, the results published on APEX-Agents 1.1 and the formalization effort requested from clients frame the promises, while some figures remain declarative.

Measured Reliability and Out-of-Lab Discrepancies

Data provided by the publisher reveals a difference between performance observed during one-off tests and reliability noted in real-world conditions. On APEX-Agents 1.1, the most effective model shows a success rate of 56.3% over a sequence of four consecutive attempts. Some agents choose ambiguous formulations to avoid strict criteria. The significance of the Series D will depend on the levels of reliability observed outside of simulations.

Required Adoption Conditions and Governance

The establishment of testing benches requires each client to draft their own business rules, necessitating a sustainable investment from internal experts before any productivity gains are realized. This process aligns with NIST recommendations regarding governance of tests prior to deployment and human oversight. Furthermore, no public data specifies the median hourly wage of the human evaluators involved.

Amount, Valuation, and Capital Movements

On November 10, Mercor raised $750 million in Series D funding, bringing its private valuation to $20 billion, a growth confirmed to have doubled since October 2025. The funding round includes Benchmark, General Catalyst, NVIDIA, Salesforce, as well as Felicis, Sapphire Ventures, and Zoom, with a stock buyback reserved for employees. Mercor is opening offices in New York and London to support major accounts and is marketing proprietary testing benches to verify business processes before production deployment. The $750 million is intended to industrialize the simulators, reflecting a shift in IT budgets towards quality control of autonomous agents.

What the Testing Infrastructure Finances

At the beginning of 2026, Mercor acquired Deeptune, whose technical teams have replicated several hundred professional applications in the form of isolated environments. These duplicates allow agents to operate without touching authentic databases, reproducing the software interfaces and internal data circuits of large companies. A second layer connects these simulators to sets of criteria as well as the APEX benchmark, designed by human professionals such as lawyers, financial analysts, or engineers. Each professional task is divided into checkpoints for binary validation to assess the compliance of the final result, and recorded failures are used to adjust models during post-training phases.

Audience, Claimed Workforce, and Caution on Figures

Mercor claims to unite over 5 million experts covering 800 professions in the United States. A post by Brendan Foody on X reached 865,000 views and 3,400 likes, indicating strong visibility in the American tech sector. However, the $20 billion valuation is a private estimate that does not correspond to actual revenue received. The figure of 5 million experts is based on declarations, without public verification of the actual number of active contributors, and the data collected on X reflects a one-time audience rather than a number of contracts concluded. The automation of office tasks relies on specialists paid to formalize their own skills, but generalist models frequently struggle to account for accounting rules and legal exceptions specific to each company.