Mercor
Mercor is a San Francisco-based AI data company that recruits domain experts, including engineers, lawyers, doctors and bankers, to train and evaluate AI models for frontier labs. It was founded in 2023 by Brendan Foody, Adarsh Hiremath and Surya Midha, high school friends from the Bay Area who left college to build it and later received Thiel Fellowships. It began as a way to match freelance software engineers with companies, then moved into supplying expert human data to AI labs. Mercor also publishes APEX, a benchmark that measures how well AI models handle professional work such as consulting, investment banking and law.
X / mercorNotable people
People map
Current peopleNotable public-facing people we cover. Grouped by public roles, not reporting lines; this is not a complete staff directory.
Leadership
- Brendan Foody
Cofounder and co-CEO
Writing, research & talks
Mercor maps 924 AI benchmarks onto 923 US occupations and 18,797 tasks, finding most working hours have little benchmark coverage and that coverage tracks occupation wage bill and AI exposure.
- Mercor Nvidia Nemotron 3 5mercor.com
Mercor post-trained NVIDIA Nemotron 3.5 Super on a few hundred APEX tasks each for legal, software and accounting work. All three domains beat the base model on held-out benchmarks, including +8.2 points Pass@1 on APEX-SWE.
- Monty Engineering Deep Divemercor.com
Mercor explains how its AI interviewer handles about 10,000 daily voice sessions: isolated per-session containers, a warm pool sized to hourly demand, failover across speech, LLM and TTS stages, and tuned turn-taking latency.
Mercor tested 12 licensed CPAs on simplified accounting benchmark tasks and compared them with frontier models. Models were faster, more accurate and far cheaper, though the tasks skew toward detail-oriented instruction following.
- How We 20x Llm Eval Throughputmercor.com
Mercor details how agent eval runs spent most of their time queued for compute, not running. A job-level scheduler replaced static per-model lanes and raised throughput several-fold from the same quota.
Mercor and Ramp introduce an accounting agent benchmark scored over repeated runs. Top models complete just over half the work, almost none succeed consistently, and results shift sharply with per-task spend budgets.
- Introducing Apex Agents 1 1mercor.com
Mercor revises its professional-services agent benchmark after finding models inflated scores by hedging across several answers. The update adds task audits, a stricter judge and prompt rules, republishes the leaderboard, and releases the dataset and reference agent.
Mercor and SkyRL walk through post-training Qwen3.5-397B-A17B with RL on expert tasks: environment and token accounting, systems tuning, an overfitting run, algorithm ablations, and evaluation. Scripts, weights and traces are released.
SWE-Marathon-Ext has eight frontier models rebuild 12 SaaS products such as Linear, Notion and Zendesk from a spec. Surface features pass often, while concurrency, time logic and UX tests expose where whole-product generation breaks down.
- Mercor Model Routing Eval Problemmercor.com
Mercor argues model routing is a Pareto frontier problem over cost and quality, measured by an agent's own eval suite, and contrasts static selection, query-level routing, and dynamic routing inside the agent.
- Agent Eval Systemsmercor.com
Mercor's framework treats agent reliability as an evaluation-systems problem. Verifiers over trajectories and outputs feed production grading, offline regressions, benchmarking and runtime gates, plus an optimization loop and data-plane deployment architecture.
- Enterprise Agent Architecturemercor.com
Mercor lays out an enterprise agent architecture split into a routing and policy plane, a runtime plane, and an intelligence plane, illustrated with a failed overnight reconciliation job inside a customer VPC.
Mercor tests whether post-training on a 2,000-task agentic dev set transfers to held-out benchmarks. Toolathalon, APEX and GDPval all improved while HLE and GPQA stayed flat, indicating real tool-use and professional-reasoning gains.
- Introducing Apex Swemercor.com
Mercor and Cognition release APEX-SWE, a benchmark of integration and observability tasks. GPT-5.3 Codex leads at 41.5% Pass@1, and an open-source harness is available for running it.
- Introducing Apex Agentsmercor.com
Mercor introduces APEX-Agents: 480 long-horizon tasks across investment banking, consulting and corporate law, built in realistic file systems and tools, with an open harness, dataset and technical report.
- Scaling Data Apex Agentsmercor.com
Mercor reports Applied Compute's post-trained GLM 4.7 now tops APEX-Agents in corporate law, with trajectory analysis: fewer steps, heavy code-execution use, and about 4x the token consumption of the base model.
- Expert Data Drives Model Performancemercor.com
Mercor and Applied Compute post-train GLM 4.6 on 874 expert tasks using long-horizon RL, nearly doubling APEX-Agents scores with no SFT warmup or filtering and a near-linear gain trend.
- Welcome To The Era Of Evalsmercor.com
Mercor's CEO argues RL can saturate any benchmark, so the limiting factor for agents is building evals that mirror real deliverables. Covers verifiable vs rubric-graded domains and why environments turn recurring work into a fixed cost.
- Expanding The Ai Productivity Index Apexmercor.com
Mercor doubles its APEX benchmark to 400 professional tasks in banking, consulting, law and medicine. GPT 5 leads at 67%, followed by Gemini 3 Pro and Grok 4, with a refined grading method and confidence intervals.
- Introducing The Ai Consumer Indexmercor.com
Mercor releases the AI Consumer Index, a rubric-graded benchmark across shopping, food, gaming and DIY. Top model GPT 5.1 scores 56.1%, and grounding failures range from 29% to 62%. Dataset and harness are open.
- Introducing Apex Ai Productivity Indexmercor.com
Mercor introduces APEX, a 200-case benchmark of investment banking, consulting, law and medical tasks built by about 100 experts. Tasks include source documents and rubrics, weighted to match real time allocation.