Google reported 68% for Gemini 4 Argon on CWE-bench v1 in its September 30 introduction. Evaluation details are incomplete.
Limits: Google reports the result; it has not been independently reproduced here. The report predates the October public API launch. The evaluated checkpoint and its equivalence to the currently served model are unknown. The task subset, sample size, harness, prompts, tools, reasoning/token budgets, run date, aggregation and confidence interval are unspecified. Model context capacity is not an evaluation budget. Evaluation safeguards and routing are unspecified. Settings reported for other models or later releases cannot fill those gaps. These are reported results, not a controlled cross-model ranking or proof of real-world reliability, speed, cost or statistically meaningful differences. Argon’s safeguard setting is unknown. Carbon’s later result is unguarded and has an unspecified benchmark revision, so the two do not establish a matched improvement.
Google
