The releases target a higher cost efficiency frontier but show varied results in external benchmarking. Artificial Analysis reports that GPT-6 Sol cut its hallucination rate from 92% to 60%, though the model declines to answer questions more often than its predecessor. While Sol's Coding Agent Index score rose to 57, Luna's score dropped 2 points to 41. Both models also exhibited regressions in knowledge work evaluations, specifically within the GDPval-AA v2.1 and AA-Briefcase v1.1 benchmarks.
