WASHINGTON—Federal research funding agencies announced a sweeping review of half a century of mathematics grants Saturday, after an artificial intelligence lab reported this week that its system had solved decades-old problems at a cost of a few hundred dollars each.

"We are not saying anyone did anything wrong," said Harold Grisby, deputy director of a grant oversight office. "We are saying two of these problems were open for more than 50 years, and the bill came to roughly what we spend on a single panel lunch."

According to a paper published Thursday in the journal Science, the system attempted 353 open problems associated with the mathematician Paul Erdős and solved nine. It checked every proof in Lean, a language that verifies each logical step. It also proved 44 of 492 open conjectures from an online catalog of number sequences.

Auditors noted that the agency's own application portal estimates 40 hours to complete a proposal, not counting the budget justification, which no one has ever completed.

Officials stressed that mathematicians remain essential, particularly for writing the 15-page section explaining why a problem matters. The system skipped that section entirely by solving the problem.

Several universities have already applied for funding to study how the system did it, proposing five years and $2.4 million. "Somebody has to understand it," said one department chair. "Ideally at our current rate."

At press time, the review committee had requested a three-year, $1.8 million extension to determine how much the review itself should cost.