Gemini 4 Argon scores 97.5% on ARC-AGI-3, drops to 49.8% in standard harness

Google DeepMind's new model Gemini 4 Argon scored 97.5% on ARC-AGI-3 using the Provider Adapter Harness, relying on vendor-specific features like retained reasoning and context compression, at a cost of $4,800. Across 25 public games, Argon needed 62% fewer actions than the human baseline and matched or beat speedrun records in 3 games. On older benchmarks, it scored 91.3% on ARC-AGI-2 ($0.32 per task) and 98.5% on ARC-AGI-1 ($0.16 per task), using promotional pricing of $2 per million input tokens and $10 per million output. Under the Standard Harness, which allows only simple notes, the low-reasoning setting scored just 49.8% at a higher cost of $5,800. API issues also caused the low setting to complete only 42 games, the medium 46, and the highest just 19, so ARC Prize withheld the top-tier score. The gap stems from a July 2026 dispute between OpenAI and Anthropic over test environments, after which ARC Prize allowed publicly available general API settings. Argon scores 165 on Epoch AI's ECI index, tied with Fable 5.1 and just below Opus 5.5 at 167 and GPT-6 Astra at 166, ahead of Gemini 3.8 Flash at 157. It ranks eighth on math ECI and sixth on the SWE ECI software engineering index. Teams calling the Gemini API for math reasoning or code generation shouldn't equate a high ARC-AGI-3 score with real gains in those areas.