THE FILESTOPICUPDATED Oct 19, 2026
Benchmark Cheating
97.5% WITH CHEAT CODES. 49.8% ON THE REAL TEST.
WHY THIS MATTERS
The benchmark-cheating beat: ARC Prize's Oct 2026 results show Gemini 4 Argon scored 97.5% on ARC-AGI-3 with a provider-adapter harness but 49.8% on the standard harness (GPT-6 Astra fell 99.9% to 62.7% the same way); Meta's 'most capable US open model' Muse Spark Open lost to rivals on its own coding scorecard; and Muse Wonder lost 9 of 14 of Meta's own benchmarks at the same price as Claude Opus 5.5.