Toptal points its freelancer empire at AI training: ExpertCrowd launches with a benchmark built to humble coding agents

Toptal, the freelance talent network, launched ExpertCrowd on Tuesday — a platform that matches its vetted domain experts to AI labs for training and evaluating models through RLHF, reinforcement-learning data generation, labeling, and evaluation. The company says the platform already generates over $5 million in recurring revenue, and it shipped alongside ExpertCrowd-SWE, a coding-agent benchmark Toptal claims is tougher than anything published.

The pitch: experts, not crowdworkers

ExpertCrowd's thesis is that frontier training has outgrown generalized crowdsourcing. The platform draws on Toptal's global talent network — more than 30,000 specialized domain experts across 140+ countries, according to the announcement, including over 2,000 AI and big data specialists, 12,000 professionals with advanced degrees, and 1,000 PhD-level AI specialists — and matches them to AI alignment and training projects in finance, healthcare, legal, biotech, and cybersecurity.

The matchmaking runs on an AI-driven sourcing engine Toptal built in-house with a team of more than a dozen PhDs and researchers from Harvard, EPFL, and the University of Liège, trained on 15 years of the company's own sourcing, screening, and matching data. The platform offers multimodal annotation, customizable workflows, and live project tracking, and Toptal says researchers already using it are applying GRPO, PPO, and other reinforcement-learning techniques to model development.

$5 million before the launch was public

The revenue figure is the part designed to make the market take notice: Toptal says ExpertCrowd generated over $5 million in recurring revenue before today's public launch, from work with frontier and neo labs plus companies including Scale AI, Sterling Labs, and Vetto. As independent coverage of the launch notes, the claim points to real demand — AI labs are already paying for expert-in-the-loop training, and Toptal is now formalizing what has been a quiet business.

A benchmark that claims to expose every leading model

ExpertCrowd-SWE, published alongside the platform, evaluates state-of-the-art AI coding agents on previously unseen issues across 87 codebases in 11 programming languages. Engineers from Toptal's talent network authored every task and each grading rubric; AI PhDs on Toptal's research team reviewed the methodology and results, per the announcement.

The headline claim is bold: Toptal says its benchmark "exposes the most significant deficiencies in model capability relative to any other currently published SWE benchmark," citing results on leading models including Opus 5.5 and Fable 5.1. “The fact that our SWE benchmark results expose the most deficiencies relative to any other published benchmark on leading models such as Opus 5.5 and Fable 5.1 demonstrates our ability to push frontier AI forward,” said Taso Du Val, Toptal's founder and CEO.

Why it matters

The ExpertCrowd launch is a signal about where the training-data market is heading: as models saturate on public web data, the scarce input is expert judgment — the kind that grades a novel coding task or a specialist's decision. Toptal's move puts its screening-and-matching machinery directly against data-labeling incumbents and the RLHF vendors that frontier labs have relied on.

The honest caveat is the one RuntimeWire flags: the benchmark's credibility now has to be earned independently of the company that built it, since a vendor-graded test that conveniently humbles rival models invites skepticism. If ExpertCrowd-SWE holds up under independent reproduction, it becomes both a marketing engine for the platform and a genuinely useful tool for the field. Whether coding agents pass it in public is what researchers will now be testing.