Snorkel says its open benchmark grants program is now a $30 million commitment

Snorkel AI says it is expanding its Open Benchmarks Grants program tenfold, from an initial $3 million to a $30 million commitment, to support researchers, domain experts and open-source teams building open evaluations of frontier AI. The announcement came in a company blog post dated October 7, 2026, written by Snorkel research fellow Vincent Sunn Chen.

The post describes three parts. The first is more funding for new benchmarks and measurement methods in areas it lists as safety and alignment, cybersecurity, physical AI, scientific agents, long-running agent work, open-ended outputs, dynamic settings and human uplift. The second is an Open Benchmarks Red Team, which Snorkel says will help benchmark builders find and fix problems such as reward-hacking exploits, contamination, broken verifiers and gaps in data diversity. Snorkel says findings will go to the benchmark team first. The third is a Snorkel Research Fellowship for independent evaluation researchers, with access to Snorkel's research team, domain experts, compute and engineering support.

The red team and fellowship are described in future terms, as programs Snorkel will be launching, and the post gives no start dates, award sizes or eligibility rules. It also does not say whether the $30 million is cash or services.

What is new is the scale, an expanded list of priority areas, and the red-team and fellowship components. The company also makes claims about its track record. It says it has worked with the teams behind benchmarks including Terminal-Bench, ARC-AGI-3 and OSWorld 2.0, and that benchmarks funded through the grants have appeared on the latest model cards from every major frontier lab. A footnote in the post names specific model cards, but this review did not check them, so both statements are Snorkel's own.

Snorkel's argument is that benchmark development has to keep pace with model development. It argues that benchmarks that become too simple, static and correlated are easy to overfit to, which distorts incentives for model builders. It calls for benchmarks that are more robust, diverse, continuously updated and independently developed.

Only one credible source, Snorkel's own post, supports this article. One other page that repeats the $30 million figure was checked and is a summary of the post. Other coverage, including a trade-press article that could not be accessed, was not reviewed, so no independent confirmation of the program's terms could be verified. The figures and the claimed track record should be read as company statements.

The announcement matters because benchmark scores are widely used to compare models, and a company whose red team will use its own data development and evaluation platform is committing to shape how that measurement is built and tested. Whether the money reaches independent teams, and on what terms, will show how much the program changes the field.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. Correction: the earlier version said Snorkel had introduced its red-team program and was offering the Research Fellowship. The source says Snorkel will be launching both programs and gives no start dates. The earlier version remains available as revision 1.