ARC Prize has officially announced that Google DeepMind's Gemini 4 Argon scored a verified 97.5% on ARC-AGI-3, making Google the second AI lab after OpenAI to break the 95% mark — behind only GPT-6 Astra at 99.9%.
ARC-AGI-3 is known as the "ultimate AGI exam": no manual, no preset instructions. The model is dropped into a completely unfamiliar grid game and must figure out the rules, discover the goal, and plan its way to clearing the level on its own. When the benchmark launched seven months ago, every frontier model scored under 1%.
Three points stand out. First, genuine learn-on-the-fly ability: across 25 public games, Argon cleared levels using 62% fewer actions than the human baseline, matching speedrun records on three games. Second, scores depend on the test setup — on the uniform standard harness, the same model scored only 49.8%, with nearly half the gap coming from the test harness. Third, older versions of the benchmark look strong too: 91.3% on ARC-AGI-2 and 98.5% on ARC-AGI-1. The Epoch Capabilities Index gives Argon 165 points, near the top, though math and software engineering remain weak spots.
Source link: Read the original. This article is a summarized digest for reference only.
