Mercor

Mercor is a San Francisco-based company that recruits professionals as contractors to train and evaluate AI models for frontier AI laboratories and enterprises. It describes itself as sitting "at the intersection of labor markets and AI research," connecting "human expertise with leading AI labs and enterprises" (Source: mercor.com). It also builds and publishes benchmarks of professional work, most prominently the AI Productivity Index (APEX) family, including APEX-Agents (APEX-Agents (Vidgen et al., Mercor, 2026)).

Snapshot

Valuation

Whether the reported $20 billion round closed has not been confirmed in the sources above.

Revenue

Business

Mercor's contractors are recruited for professional expertise rather than general annotation. Its CEO Brendan Foody has described the work as experts training models "in the same way teachers train students," with examples including doctors shaping how AI recognizes early warning signs, bankers training agents on financial analysis, and lawyers refining how models reason through precedent; he has called it "a new category of work" in which people "teach agents how to do it once, so the agent can do it a million times" (Source: mercor.com). Coverage of the company in 2025 reported that it was paying more than $1.5 million a day to people training AI systems (Source: finance.yahoo.com). Its business belongs to the contractor-vendor model described in Data-labeling and data-collection labor, alongside Scale AI.

Benchmarks

Mercor publishes the APEX family of professional-services benchmarks. The first AI Productivity Index (September 2025) covered investment banking, management consulting, "big law" and primary-care medicine; APEX-Agents (January 2026) tests long-horizon agent work in simulated project workspaces, with task design informed by a survey of 227 professionals from Mercor's marketplace and all paper authors listed with a Mercor affiliation; APEX-Agents 1.1 (September 8, 2026) revised grading after Mercor found models gaining points by "scattergunning" hedged answers. Related benchmarks include APEX-SWE (built with Cognition) and APEX-Accounting (built with Ramp), and Mercor publishes extended versions of other groups' benchmarks such as GDPval with held-out tasks of its own (APEX-Agents; APEX-Agents (Vidgen et al., Mercor, 2026)). Developers cite Mercor-run APEX results in their own model documentation; xAI's Grok 4.5 and Grok 4.6 model cards report APEX-SWE and APEX-Agents scores "conducted by Mercor" for Grok and for competing models (Grok 4.5 Model Card (xAI, July 2026); Grok 4.6 Model Card (xAI, August 2026)). A February 2026 Epoch AI review compared APEX-Agents with the Remote Labor Index and GDPval as benchmarks of economically valuable digital work, noting that APEX-Agents' environments are messier than the other two, which "probably helps with realism" and that its task-creator pay was not disclosed (What do 'economic value' benchmarks tell us? (Brand and Burnham, Epoch AI, 2026)).

Contractor conduct and quality control

404 Media reported on September 22, 2026, from internal documents and interviews with three contractors, that multiple contractors hired to improve OpenAI's models had been fired or offboarded for using AI to do the work; one internal document indicated that individual projects can involve more than ten thousand contractors. Two of the contractors worked for Mercor. Guidance to reviewers told them not to use AI or AI-detection tools and not to tell evaluators why they suspected AI use. A Mercor spokesperson said its contracts "strictly prohibit the use of LLMs to complete projects," that it invests in tools to detect misuse, and that "when we confirm an expert has used AI to complete a task, we immediately remove them from the project." OpenAI declined to comment. The same report quoted a fourth contractor, who had worked for various AI companies, saying they sometimes deliberately chose the worst responses (Source: 404media.co). The report followed 404 Media's September 14 account of OpenAI's "Project Lily," in which contractors read real ChatGPT users' prompts (AI and Privacy). The reliance of human-feedback pipelines on genuinely human output connects to concerns about Model Collapse raised in the same report.

Contractors working to improve OpenAI's models, including two who worked through Mercor, have been fired for using AI in that work, according to internal documents and interviews with three contractors that became public on September 22, 2026; one internal document indicates individual projects can involve more than ten thousand contractors (Source: 404media.co).

Policy positions

Mercor is listed among the signatories of the 2026 collective cyber-defense letter (A call for collective action on cyber defense (open letter, August 27, 2026)).

Relationships

• instance-of: Data-labeling and data-collection labor — expert-contractor variant of the annotation-vendor model
• related: Scale AI — the other large training-data vendor
• related: APEX-Agents, APEX-Agents (Vidgen et al., Mercor, 2026) — Mercor's benchmark family
• related: GDPval — Mercor publishes an extended, partly held-out version
• related: OpenAI — contractor work on OpenAI projects reported by 404 Media
• related: RLHF (Reinforcement Learning from Human Feedback), Model Collapse