One Epoch researcher noted this finally explains why systems felt slow at proofs while quietly cracking open problems: mathematics was actually cooked months ago.
Development
An AI audit finds errors in 42% of a benchmark's problems
Claim
Epoch AI shipped a v2 of FrontierMath after an AI audit corrected errors in 42% of problems, and scores leapt across the board, with GPT-5.5-xhigh reaching 85% on Tiers 1-3 and its Tier 4 result jumping from 35% to 73%, while the now-disabled Fable 5 had already topped the board at 88% on Tier 4.
Reported metric: 42% of problems wrong; Tier 4 35% to 73%
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to June 13, 2026 (theinnermostloop.substack.com)
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.