Gemini 4 Argon passes 95% on ARC-AGI-3, second after OpenAI

Google Deepmind's Gemini 4 Argon arrives at a moment when benchmark scores are increasingly contested. The model's 97.5 percent on ARC-AGI-3 came using the Provider Adapter Harness, which relies on vendor-specific features like retained reasoning and context compaction. Under the standard testing harness, the score falls to 49.8 percent, a gap that traces back to a July 2026 dispute over how different API settings affect results. That context matters because it shows how much benchmark outcomes can shift based on the testing setup rather than the model alone.

Epoch AI's tracking places Argon at 165 points on its Epoch Capabilities Index, tying Fable 5.1 and sitting just behind Opus 5.5 (167) and GPT-6 Astra (166). This is the closest Google has been to the frontier since Gemini 3.1 Pro in February 2026, well ahead of Gemini 3.8 Flash (157 points). But the gains are uneven: early estimates rank Argon eighth on the Math ECI and sixth on the Software Engineering Index, suggesting coding and math remain weaker spots even as knowledge, agent, and vision benchmarks improve.