Overview
- Google disclosed Wednesday that it is rolling Gemini 4 Argon to a small group of vetted cybersecurity partners and internal teams through its Fairwind program while broader access is paused for further testing.
- Argon raises the model output cap to one million tokens so it can produce much longer single‑trajectory results for tasks like long reports, multi‑stage code changes, and sustained agent workflows.
- Google published benchmark results showing Argon leads or ties top rivals on many domain tests, including a 77.9% on DeepSWE v1.1 and top marks on several knowledge‑work and long‑video benchmarks, though it trails some competitors on other coding tests.
- Inside Google the model is already used for real tasks such as fleet memory optimizations that freed about 300 TiB and large C/C++ to Rust code migrations, and the company says it trained Argon to support advanced cybersecurity defense and vulnerability finding.
- Before wider release Google says it will harden guardrails by monitoring internal activations, improving prompt‑injection defenses, isolating sandboxes, participating in the U.S. voluntary pre‑release review, and collecting feedback from early testers; introductory pricing has been reported but final commercial terms are pending.