Overview
- Google announced Gemini 4 Argon on Wednesday, Sept. 30, 2026, and immediately gave the model to a vetted set of cybersecurity partners and internal teams for testing.
- Argon raises the model output cap to one million tokens so it can sustain very long, single‑trajectory reasoning for tasks in software engineering, legal and financial research, and complex enterprise workflows.
- Google published benchmarks showing Argon leading or tying top rivals on many knowledge‑work and engineering tests while noting mixed results on some coding benchmarks where competitors scored higher.
- The company says it trained Argon for cybersecurity defense and will provide it without cyber guardrails to trusted defenders, and it is participating in the U.S. voluntary pre‑release safety process while hardening prompt‑injection defenses, sandboxing, activation monitoring, and misalignment checks.
- Google disclosed introductory API pricing and a staged commercial rollout that begins with paid API customers and AI Ultra subscribers, and it highlighted internal wins such as using Argon agents to free hundreds of tebibytes of memory in Google data centers.