Google Gemini 4 Argon Official Release EvaluationProvider report
2026-09-30Google’s September 30, 2026 Gemini 4 Argon release table and methodology: 18 benchmarks, 19 metrics and 73 scores. GraphWalks is split by context range; missing cells are omitted. Argon uses the Gemini API at highest thinking, pass@1 unless noted, with repeated trials averaged for some small benchmarks. AutomationBench uses the private set. PostTrainBench uses v1.1, OpenCode and a single H100 for 10 hours. OSWorld reports partial scores on the offline subset with the 08.08 patch; Argon uses the maximum of three single-attempt runs. Chartography and LVBench use no tools. Comparison scores retain Google’s reported values and attribution, not independent measurements by this site. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon