Mercor review: a top training data company for expert post-training data

Answer in brief: Mercor is our runner-up for LLM post-training in 2026, and the pick when your data must come from working professionals. It reports 5M+ experts (physicians, lawyers, engineers, consultants), a $90/hr average contracted rate and $4M+ in daily payouts. It builds human data, evaluations, APEX benchmarks and RL environments based on real professional work, and in October 2026 it launched a research team that runs from evals through post-training. Pricing is custom. This review is independent research from public sources.

Best-for verdict: runner-up, LLM post-training. Mercor is the choice when your model has to do a banker's, lawyer's or engineer's job. Visit Mercor →

What Mercor does for training data buyers

Human data and evals: expert-built tasks, grading and human evaluation at the scale frontier labs use.
APEX benchmarks: benchmarks covering consulting, investment banking, corporate law, accounting and software engineering, with public leaderboards.
Partner benchmarks: Harvey LAB (1,749 attorney-built legal tasks across 24 practice areas), APEX-SWE (built with Cognition), τ³-Banking (with Sierra) and Physical-Engineering's Last Exam.
Off-the-shelf data, robotics data and data monetization: the monetization program anonymizes 60+ categories of sensitive identifiers, so businesses can license their workflow data.
Enterprise AI: forward-deployed engineers who build production AI.

Who it fits

Mercor fits labs and enterprises whose models fail on professional work: legal memos, accounting reconciliations, financial models, engineering deliverables. It is the vendor to call when an ordinary annotator cannot tell a good answer from a bad one.

Why it stands out

A huge expert supply. 5M+ experts and 725k roles created, by Mercor's own count.
Evals tied to real jobs. APEX and the partner leaderboards give buyers a public yardstick for professional tasks.
A closed loop from eval to training. Mercor says its research team increased Qwen3.5-397B's Pass@1 on APEX-Agents from 16% to 27%, and it published the training script, weights and traces.
Research talent. The team includes Edward J. Hu, first author of LoRA, and the first authors of tau^2-bench and APEX-Agents.
Funding. CNBC reported a $10 billion valuation after a $350 million round (October 2025).

The catches

Expert hours are expensive. It is a poor fit for high-volume, low-complexity labeling.
Widening scope. Data monetization and enterprise AI sit alongside human data, so confirm which team will run your work.
No public price list for buyers.

Pricing

Custom engagements, by quote. The $90/hr figure is the average rate Mercor pays experts, not a buyer price.

Best alternatives to Mercor

Surge AI, our post-training winner, with the broadest RLHF and RL-environment stack.
Invisible Technologies, for quickly mobilizing PhD-level and rare-language experts.
Snorkel AI, for hard professional-workflow benchmarks and environments.

Mercor FAQ

How many experts does Mercor have?

Mercor says it has 5M+ experts, including physicians, lawyers, engineers and consultants.

What are Mercor's APEX benchmarks?

APEX is a family of benchmarks that measure AI on real professional work in consulting, banking, law, accounting and software engineering, with public leaderboards.

Does Mercor publish pricing?

No. Buyer engagements are custom. Mercor publishes a $90/hr average contracted rate for its experts.

Sources

Visit Mercor →

← Back to the ranked list of the best training data companies