Google Gemini 4 Argon Hits 97.5% to Become Second Lab to Top ARC-AGI-3

Google DeepMind's Gemini 4 Argon reasoning model has achieved 97.5% accuracy on the ARC-AGI-3 benchmark using a provider adapter harness costing $4.8K, making it the second AI lab to exceed 95% on the test. The model also scored 91.3% on ARC-AGI-2 and 98.5% on ARC-AGI-1, with François Chollet noting similarities to Astra's performance profile. Analysts attribute the improved scores to reinforcement learning, estimating an 85% probability that RL drove the result compared to a 15% chance for 'loopers.'