Snorkel’s $3.5 Billion Bet on Model Evaluation

In September 2026, Snorkel AI raised $350 million at a $3.5 billion valuation. Five years earlier, coverage described it as a data-labeling service valued at $1 billion. By 2026, Snorkel was pitching a platform where humans and AI agents create and vet data. The shift raises a customer question its valuation cannot answer: who keeps the judgments used to make a model useful?

The same base model can produce a different product

In August 2026, Z.ai said GLM-5.3 used the same base model as GLM-5.2 but gained stronger coding skills through scaled post-training. Z.ai said it planned to release the newer model’s weights two weeks later; that announcement alone does not establish that the release happened. Its comparison isolates an engineering distinction: a base model can stay fixed while later work changes what users get from it.

Pre-training, fine-tuning and post-training change model weights for different purposes, as Chip Huyen distinguishes in AI Engineering. Evaluation does not itself change weights. It supplies the tests and judgments that tell a builder what to change next. A company adapting a model may come to depend less on any single tuned version than on the process connecting its domain data, adjustments and measured results.

OpenAI’s 2024 expansion of its Custom Model program put assisted fine-tuning, training-data pipelines, evaluation systems and training and validation metrics into one customer workflow. Organizations can buy help adapting a model while remaining dependent on its provider for the underlying service. Open weights offer a different starting point, but they do not supply a customer’s task definitions or tests.

Snorkel moved from supplying examples to organizing judgment

In August 2021, coverage reported Snorkel’s $85 million raise at that $1 billion valuation. By May 2025, Forbes described a pivot to AI evaluation tools and a reliance on human experts to judge system accuracy; it put Snorkel’s valuation at $1.3 billion. In September 2026, Reuters described an “agentic data development platform” on which humans and AI agents create and vet data, alongside the $350 million raise.

Snorkel’s three reported valuations track a changing pitch, but they do not measure customer adoption or show that evaluation displaced labeling revenue. They cannot establish a market-wide adoption curve.

Other builders attacked adjacent parts of the problem independently. Cleanlab was selling tools for more accurate training data in 2023, before Snorkel’s reported evaluation pivot. Cleanlab’s data-quality work and Snorkel’s expert-evaluation emphasis are different products. Both address a constraint for model builders: more examples help less when they cannot tell which examples or outputs deserve trust.

Evaluation becomes part of operating the model

AWS introduced Model Evaluation on Bedrock in preview in 2023 to help companies assess models before use, with human testers involved. Snorkel’s later emphasis on expert judges addresses a harder question than whether testing is available: who decides which mistakes matter? Someone must translate an operational goal into acceptable answers, consequential failures and cases a model must pass. A fluent answer and a correct answer can receive very different marks.

A pre-deployment score has a shelf life. Huyen’s Designing Machine Learning Systems describes how changes in an application’s environment or users can produce covariate shift. The organization using the model must then check whether its earlier test cases still represent the work. To maintain operational assurance, the customer needs to retain test cases, review outputs and revise the rubric after launch.

Meta’s Scale stake drew inquiries, not a buyer exodus

Meta’s $14.3 billion investment in Scale AI exposed the commercial sensitivity of this work. In June 2025, Snorkel, Labelbox and Turing reported increased client interest from customers concerned about Scale’s independence. A vendor helping produce proprietary training data and judgments may learn what a customer’s model gets wrong and how it intends to fix those errors. Reported interest shows a procurement concern, not completed migrations to Scale’s rivals.

Scale still won large contracts. The Department of Defense awarded it a $500 million contract in 2026 to help sift through data and assist decisions, after a $100 million deal in 2025. The Pentagon’s choice cautions against treating inquiries to Scale’s rivals as a wholesale switch.

A customer’s standards need an exit path

A buyer can put domain language and business constraints into training examples and evaluation cases rather than asking a generic model to infer them. It must still decide who can inspect those examples, revise labels, challenge an expert judgment and carry validation histories to another provider. The procurement test is whether the customer can retain and move those artifacts, not merely whether a vendor offers an evaluation dashboard.

Human review also has a cost that agentic tooling cannot wish away. Huyen notes that bias can enter through data collection, sampling and labeling. Wired reported in 2023 that some AI data-labeling providers had inadvertently hired minors, sometimes exposing them to traumatic content. A buyer relying on human feedback needs to know how that work is sourced and supervised, as well as how the resulting judgments affect model behavior.

Snorkel’s funding rounds and AWS’s evaluation launch show suppliers building this layer, but say little about how many customers can carry their work between providers. Snorkel’s $3.5 billion valuation cannot tell a buyer whether its examples, rubrics and validation histories will follow it to the next vendor. If they cannot, the provider that makes customization easiest also keeps the definition of “correct.”