Google announces Gemini 4 Argon, phased to Fairwind cyber defenders first

Google on 30 September 2026 announced Gemini 4 Argon, its new frontier model, in a blog post by Koray Kavukcuoglu, SVP at Google DeepMind and Google's chief AI architect. The model is not broadly public yet. Google says it is rolling out first to trusted cyber defenders through the Fairwind Program, with a phased path that includes the U.S. government's voluntary pre-release access process.

Google pitches Argon for complex software engineering, legal and finance knowledge work, and cybersecurity defense. Vendor-reported pricing for the introductory period is $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off. Google also says the output token limit rises to 1 million tokens from 64K. After the intro period, Google lists $4 / $20 per million input / output tokens.

Internal Google examples in the same post (quantum subroutine optimization beating a published baseline by 40%, large memory savings from agent-driven telemetry work, and C/C++ to Rust migrations such as parts of libgav1 and Fuchsia Zircon) are Google-reported, not independent audits. Google also cites early Fairwind use by Wiz that found a critical flaw in hospital software other models had missed.

The Guardian reports that access is restricted to vetted cybersecurity experts over misuse fears, compares the approach to Anthropic's restricted Claude Mythos Preview, and places the timing next to a White House voluntary accord on AI risk policing. Google lists safeguards including misuse and CBRN refusal, stronger prompt-injection robustness on Gray Swan's IPI bench, misalignment monitoring of chain-of-thought, and hardened sandboxes. Broader availability is described as coming first to paid API customers and Google AI Ultra subscribers.

Key facts

• Primary: Google Gemini 4 Argon post (Koray Kavukcuoglu), 30 September 2026
• Independent: The Guardian, 1 October 2026
• Access path: Fairwind Program cyber defenders first; U.S. voluntary pre-release process
• Intro price (Google-reported): $2/M input, $10/M output; cached input 95% off
• Output limit (Google-reported): 1M tokens, up from 64K
• Next: paid API customers and Google AI Ultra, per Google

How to read it

This is a staged vendor launch, not a public API drop. Bench scores, internal productivity examples, and the hospital vulnerability find are Google-reported. The Guardian's framing stresses the restriction and the Anthropic parallel. Treat security and capability claims as vendor claims until outside teams publish their own results.

Independence

Frontier Flux is not affiliated with Google, DeepMind, The Guardian, Wiz, Gray Swan, or Anthropic. Figures and quotes come from Google's post, the Fairwind program page, and the Guardian report, not from our testing.