Gemini 4 Argon: 97.5% or 49.8% on ARC-AGI-3
Why it matters
ARC Prize verified Google's Gemini 4 Argon at 97.5% on ARC-AGI-3 (low reasoning, Provider Adapter harness, about $4.8K), second only to GPT-6 Astra's 99.9%, with 62% fewer actions than the human baseline on the 25 public games. With the standard harness, which only carries forward notes the model keeps, it scored 49.8% (low) and 46.3% (medium). Caveat: API issues left the standard-harness runs incomplete (42 of 55 games at low reasoning), with partial credit counted. Near the ceiling, the harness can change a score by about half.
Read the source: ARC Prize (opens in a new tab)
All hits from 19 Oct 2026More ResearchGet the daily hits by email