Google DeepMind’s Gemini 4 Argon scored 97.5% on the ARC-AGI-3 reasoning benchmark when tested through a provider-specific setup that preserved reasoning and compacted context. That made Google the second lab after OpenAI to exceed 95% under the ARC Prize’s expanded testing rules.
The result changed substantially in the Standard Harness, which permits only simple notes between rounds. At low reasoning effort, Argon scored 49.8% and cost about $5,800 for the run, compared with roughly $4,800 for the higher result. API problems also meant it did not complete all 55 games in the standard environment, so the published figures are not a clean like-for-like comparison.
On the 25 public games, Argon and OpenAI’s GPT-6 Astra each used 62% fewer actions than the human baseline. Argon also recorded 91.3% on ARC-AGI-2 and 98.5% on ARC-AGI-1 at introductory API prices. Separate evaluations cited by The Decoder found it behind rivals on some mathematics and software-engineering tasks.
The gap is the practical headline: a benchmark can measure the surrounding API and memory mechanism as well as the model. Argon’s strongest score is notable, but users should not treat it as a context-free measure of general intelligence or expect identical performance through every interface.