Toptal, the freelance talent marketplace, is going after the AI training-data business. The company on October 14 launched ExpertCrowd, a platform that matches AI developers with vetted domain experts to generate the RLHF data, reinforcement-learning datasets, and model evaluations that frontier labs increasingly depend on.
The pitch is specialization. While data-labeling work has traditionally flowed to general crowdsourcing platforms, Toptal argues that building more capable models requires expert judgment -- a cardiologist, a securities lawyer, or a PhD chemist -- applied across industries like financial services, healthcare, legal, biotechnology, and cybersecurity.
30,000 experts, on demand
ExpertCrowd is built on Toptal's existing global network, which the company says spans more than 30,000 specialized domain experts across over 140 countries, including more than 2,000 AI and big-data specialists, 12,000 professionals with advanced degrees, and 1,000 PhD-level AI specialists available on demand.
Matching happens through an AI-driven sourcing engine developed in-house by Toptal's AI research team, made up of more than a dozen PhDs and researchers from institutions including Harvard, EPFL, and the University of Liege, and trained on 15 years of the company's proprietary data in sourcing, screening, and matching skilled experts.
The platform offers multimodal annotation, customizable workflows, and live project tracking. It is aimed at frontier AI labs, enterprises, and startups, and Toptal says researchers applying GRPO, PPO, and other reinforcement-learning techniques have already used it in developing their models and systems.
Already past $5M in recurring revenue
ExpertCrowd arrives with traction. Toptal says the platform generated more than $5 million in recurring revenue before its official launch, from work with frontier and newer AI labs, as well as companies including Scale AI, Sterling Labs, Vetto, and other startups that service the frontier labs.
That customer list is telling: Scale AI built one of the largest data-labeling businesses in the industry, and its appearance as an ExpertCrowd customer suggests demand for expert-level data now exceeds what any single provider can supply
The launch also sharpens a growing divide in the AI data market -- between cheap, commoditized annotation and high-margin expert workflows. As models get smarter, the bottlenecks increasingly sit in domains where only specialists can judge correctness, and pricing is following them there.
A new, tougher benchmark for coding agents
Alongside the platform, Toptal released ExpertCrowd-SWE, a benchmark designed to evaluate leading AI coding agents on previously unseen issues.
The benchmark spans 87 codebases across 11 programming languages, with every task and grading rubric authored by engineers from Toptal's talent network and reviewed by PhD-level AI researchers.
Toptal's claim is bold: its benchmark exposes the most significant deficiencies in model capability relative to any other currently published software-engineering benchmark. It names Opus 5.5 and Fable 5.1 among the leading models tested.
"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," founder and CEO Taso Du Val said.
With labs racing to sell autonomous coding agents, independent benchmarks that genuinely stress-test them are becoming strategic assets -- and Toptal is betting that owning the talent pipeline behind both training and evaluation is an even better one.
By Hamza Faisal
