Google DeepMind reports Gemini 4 Argon scores 97.5% on ARC-AGI-3 with specific harness, drops to 49.8% in standard test

Google DeepMind's new frontier model, Gemini 4 Argon, has shown significant progress on the ARC-AGI-3 benchmark, though its performance depends heavily on the testing method. According to The Decoder, the model scored 97.5% at a cost of $4,800 using the "Provider Adapter Harness," a setup that allows vendor-specific features like retained reasoning and context compaction. This makes Google the second AI lab after OpenAI to cross the 95% mark on what is widely considered the toughest benchmark for general problem-solving. However, in the "Standard Harness," which only allows simple notes between rounds, the score drops sharply to 49.8%.

Research institute Epoch AI tracked Google's progress, noting that Gemini 4 Argon scored 165 points on the Epoch Capabilities Index (ECI), tying Fable 5.1 and landing just behind Anthropic's Opus 5.5 (167) and OpenAI's GPT-6 Astra (166). This marks the closest Google has been to the current AI frontier since February 2026. While leading in several knowledge, agent, and vision benchmarks, early estimates place it eighth on the Math ECI and sixth on the Software Engineering Index (SWE ECI).