GPT-5.5 "9.7T parameters" re-evaluated: after correction, only about 1.5T remain

CryptoWorld News reports that AI researchers Lawrence Chan and Benno Sturgeon have rechecked a paper by Pine AI Chief Scientist Li Bojie titled “Incompressible Knowledge Probes: Estimating the Parameter Count of Black-Box Large Language Models Based on Factual Capacity.” The original paper estimated GPT-5.5 at approximately 9.7T, Claude Opus 4.7 at about 4.0T, and O1 at around 3.5T. The recheck found that the original numbers were inflated due to the scoring methodology and question quality—especially due to improper handling of “floor scores,” which exaggerated the performance gap on difficult questions. After removing “floor scores,” the fitted slope fell from 6.79 to 3.56, r² dropped from 0.917 to 0.815, and the 90% prediction interval widened from 3.0 times to 5.7 times, indicating larger errors. The recheck also found that 131 questions had ambiguities or incorrect answers, accounting for 9.4%. Based on the corrected data, the parameter count for GPT-5.5 decreased from 9659B to 1458B, and the 90% prediction interval was 256B to 8311B. The recheck authors emphasized that 1.5T should not be taken as the true parameter count of GPT-5.5.