Independent tests rank Gemini 4 Argon first on Vals Index, with the lowest hallucination rate among top models

The Batch reviewed independent benchmark results for Google's Gemini 4 Argon on 9 October 2026. On Vals AI's Vals Index, a mix of eight finance, coding, legal and tax tests, Argon at high reasoning ranked first of 43 models with 68.9% accuracy. On AA-Omniscience, a factual knowledge test that penalises wrong answers but not refusals, it had a hallucination rate of about 15%, the lowest among models scoring 45 or more on Artificial Analysis' Intelligence Index, against about 51% for GPT-6 Astra at max reasoning. It is not best at everything. The Batch says Argon ties for third on Artificial Analysis' overall Intelligence Index, matching GPT-6 Astra and Claude Fable 5.1 at lower cost per task, and still trails Anthropic's newest models on agentic coding. Argon is currently available only to organisations in Google's Fairwind cybersecurity programme.