Surge AI vs Mercor: which expert data provider is better for AI labs in 2026?
Answer in brief: Surge AI is the pick when you want a data partner that proves its data on real models: its Training Lab publishes post-training gains, and Trusted Program labs can pay for off-the-shelf data only if it works. Mercor is the pick when you need professional depth at volume. It says it has 5M+ experts, including physicians, lawyers, engineers and consultants, plus a large licensable catalog with same-day samples and benchmarks built with partners such as Harvey, Cognition and Sierra.
Top pick for proven post-training lift: Surge AI. Its public before-and-after results and train-first terms are the strongest risk reducers in the market.
This is independent research from published vendor material and current reporting, not hands-on testing. See our methodology.
The head-to-head table
| Factor | Surge AI | Mercor |
|---|---|---|
| Model | Managed expert data company with its own Training Lab | Expert marketplace and human-data company |
| Expert pool (company claim) | Domain experts across STEM and the humanities; 70+ languages | 5M+ experts; $90/hr average contracted rate |
| Off-the-shelf data | Coding agents, enterprise agents, multimodal and expert reasoning; pay-if-it-works for Trusted Program labs | 50K+ tasks across 30+ domains; samples delivered the same day |
| Example datasets | Long-Horizon Tool-Use, Frontier Work, Chartography, ComplexConstraints | APEX-Agents (2,620 tasks), SWEBench Ext. (12,000), TerminalBench (10,000), BrowseComp (3,558) |
| Public evidence | Self-reported model gains in the Training Lab | Partner benchmarks: Harvey LAB, APEX-SWE (with Cognition), τ³-Banking (Sierra), APEX-Accounting |
| Other business lines | Focused on data | Enterprise AI deployment, data monetization, robotics |
| Pricing | Custom; off-the-shelf data on pay-if-it-works terms | Custom; datasets licensed after sampling |
Surge AI: data with receipts
Surge's pitch is evidence. On its Training Lab page it reports that training Kimi K2.7 on 2,400 expert-built coding tasks raised DeepSWE from 31.0% to 48.7%, and that training on office-work data with no coding tasks still improved Terminal-Bench 2.1 by 8.4 points. Surge's product line covers RL environments and agents, rubrics and verifiers, RLHF, SFT and human evaluation. These are exactly the post-training levers a lab pulls. The gains are self-reported, so confirm them on your own evals.
Mercor: the deepest professional bench
Mercor's strength is the size and seniority of its network. Its home page lists a $90/hr average contracted rate and $4M+ in daily payouts. Its off-the-shelf catalog says every task is written by PhD researchers, practicing lawyers and senior engineers, is peer reviewed, can be licensed for commercial training and can be extended through custom projects. Its partner benchmarks include Harvey LAB, an attorney-built set of 1,749 legal tasks across 24 practice areas. They show Mercor can turn professional work into a gradeable signal. CNBC reported a $10B valuation after a $350M round in October 2025.
Same battlefield: professional knowledge work
The two suppliers increasingly meet on the same ground: making models good at real office deliverables. Mercor's APEX-Agents dataset runs 2,620 expert-built tasks inside enterprise app clones across eight domains, led by finance (830 tasks), consulting (590) and law (530). Mercor says its datasets support SFT, DPO and GRPO training. In October 2026 it introduced a research team that measures AI on real work, builds expert data around where models fail and post-trains to close the gap.
Surge attacks the same problem with its Frontier Work and DAYJOB-style data. It reports that post-training GLM-5.3 on professional-work data raised its GDPval-AA Elo rating from 1650 to 1704, and that Kimi K2.7's GDPval success more than doubled after training on 1,465 GDP.pdf tasks. Both sets of figures are vendor claims. The practical test is simple: give both suppliers the same failing professional tasks from your evals and compare the lift.
Pick Surge AI if…
- You want spend tied to measured model improvement.
- Your priorities are agentic coding, long-horizon tool use, instruction following or multimodal document reasoning.
- You prefer a supplier focused only on data.
Pick Mercor if…
- You need licensed professionals (law, accounting, medicine, finance, engineering) at high volume.
- You want to license a large catalog quickly and review samples the same day.
- You want benchmarks built with industry partners to anchor your evals.
Pricing model
Both companies quote custom work. Surge is the only one of the two with a published pay-if-it-works term, which applies to Trusted Program labs. Mercor publishes what experts earn (a $90/hr average contracted rate), not what buyers pay. Expect a markup on top of that, and ask for a per-task quote.
FAQ
Which is bigger, Surge AI or Mercor?
Mercor reports the larger network, with 5M+ experts. Both are private, so revenue figures are unaudited. Size matters less than whether the right experts are available for your task.
Do Surge AI and Mercor sell off-the-shelf datasets?
Yes. Surge offers off-the-shelf training data, evals and RL environments. Mercor licenses datasets such as APEX-Agents, SWEBench Ext. and TerminalBench, with same-day samples.
Does Mercor do more than training data?
Yes. Mercor also sells enterprise AI deployment with forward deployed engineers, a data monetization program and robotics data. Confirm which team will staff your project.
Keep comparing
·
Reviews: Surge AI review · Mercor review. More: the ranked list, alternatives to Mercor, alternatives to Surge AI, Mercor vs Handshake AI. For dataset marketplaces, see Best AI datasets.
Sources
- Surge AI Training Lab
- Surge AI off-the-shelf data
- Surge AI products
- Mercor home page
- Mercor off-the-shelf data
- Mercor: why we are building a research team
- CNBC: Mercor valued at $10 billion
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.