Google has announced Gemini 4 Argon, its new frontier AI model, and almost nobody can use it yet. Initial access goes to a vetted set of cyber defenders through what Google calls its Fairwind Program. Google’s own internal teams also get it. Both groups get a build without its cyber guardrails, the limits meant to stop misuse for hacking. Google frames a broad public launch as contingent on more safety work and testing, without giving a date.
Who Gets Gemini 4 Argon First
The Fairwind Program is the channel for that early access. Google’s own announcement spells out the arrangement:
For trusted defenders and our own internal teams at Google, we’ll be releasing Argon without cyber guardrails so they can leverage its full frontier-level cybersecurity defense capabilities.
Separately, Google says it has built in checks against misalignment. Those checks watch what the model reasons through and what it does, and they can halt a task mid-run. Google says a comparable setup flagged problems during its own training runs to an internal response team. These are two different things and worth keeping apart. The guardrail-free build is a scoped exception for vetted defenders, not a preview of how a public version would ship.
What Google Says Argon Can Do
Google positions Argon as a model for deep reasoning across long, multi-step work. The biggest concrete change for developers is the output ceiling. Argon can return up to 1 million tokens in a single response, against 64,000 on the previous model. Google’s published benchmark results include the following.
- DeepSWE v1.1 — a score of 77.9% on the software-engineering benchmark
- CWE-bench v1 — 68%, which Google says ties for first place
- AutomationBench — 51.3%, which Google says is the top score on Zapier’s benchmark
- LVBench — 91.7% on long-video understanding, which Google calls state of the art
Google also lists results from its own engineering work. It says agents built on Argon freed over 300 TiB of data-center memory. The same effort moved C and C++ codebases over to Rust, including more than 800,000 lines of Fuchsia’s Zircon kernel. It also rewrote the libgav1 video decoder to run 2.7 times faster. These are Google’s own numbers, measured on Google’s own infrastructure.
Pricing and Availability
Argon is rolling out to trusted cyber defenders now through Fairwind. When access widens, Google says it starts with Google AI Ultra subscribers and paid API customers. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95%. After the introductory period, those rates rise to $4 and $20.
That is a doubling once the introductory window closes. Anyone costing out a project at launch rates should treat them as temporary rather than real. A single response that fills the 1 million token ceiling runs $10 in output charges today, and $20 later.
For most people reading this, Argon is not something to try today. It matters mainly as context. Gemini models already power a lot of what Android and Google’s apps do.
Source: Google
