Snorkel AI iconSnorkel AI

Snorkel AI sells expert-built training data, RL environments, evaluation rubrics and benchmarks to frontier labs and AI teams as data-as-a-service

commercial enterprise
$350M Series E (Sep 2026)
$3.5B Valuation
2019 Founded (Stanford AI Lab spinout)

Overview

Snorkel AI says "We build the data that pushes the frontier": it "helps frontier labs and AI teams develop specialized training data and environments" for their models and agents. That is a big change from its origins. Snorkel was founded in 2019 by Stanford AI Lab researchers, and its first product, Snorkel Flow, was a platform that automated labeled-dataset creation with programmatic labeling. Snorkel has since moved from selling software to selling ready-made data: reinforcement learning prompt sets written by tens of thousands of human experts, evaluation rubrics refined with reviewer feedback, and training sandboxes such as simulated developer workstations for coding models. It also publishes benchmarks (Terminal-Bench, Senior SWE-bench, OSWorld 2.0, Agents' Last Exam and others), vertical datasets (SnorkelFinance, SnorkelLegal, SnorkelUnderwrite) and runs an Open Benchmarks Grants program. Snorkel Flow no longer appears on the homepage, and its old product URL now describes the data services. In September 2026 Snorkel raised $350M at a $3.5B valuation, co-led by Insight Partners and S32; CEO Alex Ratner cited an annualized revenue run rate of $375M (company claim).

The Verdict

Who Should Use Snorkel AI?

Best For

  • Frontier labs buying expert RL and supervised training data
  • AI teams that need custom evaluation rubrics, benchmarks or agent sandboxes
  • Enterprises in finance, legal or insurance that want domain datasets

Not Ideal For

  • Teams that want self-serve labeling software (see Label Studio or Labelbox)
  • Small teams that need public, pay-as-you-go pricing
  • Anyone looking for the original open-source Snorkel library as a maintained product

What's Great

  • Expert-written data for RL, evaluation and agent training, not just labels
  • Strong research roots (Stanford AI Lab) and public benchmark work
  • Well funded after the September 2026 round

Watch Out For

  • No public pricing (snorkel.ai/pricing returned 404 in November 2026)
  • Snorkel Flow software is no longer marketed on the site
  • Cut 13% of staff in September 2025 while pivoting to data-as-a-service
  • Very few public product reviews to judge quality

Pricing

View all features & details

What Snorkel Sells

  • Expert-curated datasets for supervised and reinforcement learning
  • Evaluation environments, rubrics, reviewer guidance, difficulty tiers
  • Training sandboxes, such as simulated developer workstations
  • Custom data development for frontier models

Benchmarks and Datasets

  • Agents' Last Exam, OSWorld 2.0, Terminal-Bench, Senior SWE-bench, SlopCode Bench
  • SnorkelFinance 2.0, SnorkelLegal, SnorkelUnderwrite 2.0, MBABench 2.0
  • Open-source snorkel library (Apache-2.0) still on GitHub; its README says the team now focuses on Snorkel Flow

Real-World Usage

Who It Works With

  • Logos on the homepage include Google, AWS, Microsoft, Anthropic, Mistral AI, OpenAI and Stanford University (logo wall, not confirmed customer list)

Growth

  • CEO: grew more than 18x since launching data-as-a-service about a year earlier
  • Annualized revenue run rate of $375M (company claim)

How It Compares

Feature Snorkel AI Labelbox Encord Label Studio
What you buy Finished expert datasets and environments Platform plus expert services Data platform Open-source software plus cloud
Software you run yourself Not marketed Yes (free tier) Yes, sales-led plans Yes, open source
Public pricing No In docs No Yes
Latest funding $350M Series E (2026) $100M+ Series E (2026) $60M Series C (2026) Not covered here

Sources: snorkel.ai · Labelbox Series E · Encord Series C · HumanSignal pricing · snorkel library

Is this your tool? Add a badge