Google’s Gemini 4 Argon: 1M output tokens, Fairwind first, no public date
Koray Kavukcuoglu, 30 September: trusted cyber defenders first, US voluntary pre-release, then paid API and AI Ultra. DeepSWE 77.9%, Vals Index lead, $2/$10 intro. Ars: so much for Gemini 3.5 Pro.
Deep dive3 minIn-depth
Maya Chen3 min read
- Gemini
- Argon
- Fairwind
- Cybersecurity

Photo: GoogleSource page
Google has a new frontier model and a short list of people who can use it. Koray Kavukcuoglu, SVP at Google DeepMind and the company’s chief AI architect, posted Gemini 4 Argon on 30 September. It is rolling first to trusted cyber defenders in the Fairwind Program. Google is in the US government’s voluntary process for pre-release model access. Broad availability, he wrote, will start with paid API customers and Google AI Ultra subscribers. There is no public date. Lucas Ropek at TechCrunch, 4:43 p.m. Pacific, called it Alphabet’s most powerful model yet and noted the cyber-first gate. Ryan Whitwam at Ars, 4:11 p.m.: you cannot use it yet.
The output window is 1 million tokens, up from 64,000. Kavukcuoglu’s claim is that the extra headroom lets the model finish long jobs in one trajectory. Introductory API price: $2 per million input tokens and $10 per million output, with cached input at 95 percent off. After the intro period, $4 and $20. On DeepSWE v1.1, which Google uses for long-horizon software engineering, Argon scores 77.9 percent — a state-of-the-art the company claims for that bench. It leads the Vals Index, which weights finance, coding, legal and tax work by US GDP. AutomationBench, Zapier’s end-to-end business-function test: 51.3 percent, first. LVBench, long video: 91.7 percent. CWE-bench v1, remediating vulnerabilities: 68 percent, tied for first. Whitwam printed Google’s DeepSWE comparison against GPT-6 Astra, Fable 5.1 and Opus 5.5. Those rows are Google’s.
Inside Google, Argon is already on the job. A team of agents read fleet-wide profiling telemetry and applied memory fixes that, once rolled out, freed more than 300 tebibytes, with 500 tebibytes to 1 pebibyte estimated later. Other agents are moving C and C++ to Rust: tens of thousands of lines in re2 and libgav1, more than 800,000 in the Fuchsia Zircon kernel. Kavukcuoglu says those rewrites are under automated and manual audit, emulation and review before production. For libgav1, agents replaced 32,000 lines of SIMD in an existing Rust port; the result, Google says, is a memory-safe decoder 2.7 times faster than that port, with identical video output. Quantum researchers used Argon to cut spacetime resources on a bottleneck subroutine 40 percent below a published baseline, in minutes.
Built to sustain deep reasoning across complex, long-horizon workflows, Argon is fundamentally changing the way we work and build at Google.
— Koray Kavukcuoglu, Google, Sep 30, 2026
Fairwind partners get Argon without cyber guardrails. Wiz is already running it on Scan for Good, which looks for high-risk exposures in public infrastructure at no charge. Google says the model found a critical flaw exposing personal data in healthcare software used by hospitals worldwide, one that earlier frontier models missed. It did not name the product. On Google’s internal vulnerability bench, Argon found exposures across codebases in 20 languages. On Wiz’s black-box web test, Google says it beat 3.8 Flash Cyber at mapping the attack surface, finding bugs and writing proof-of-concept evidence. Safeguards listed for the later public release: Frontier Safety Framework refusals for cyber and CBRN misuse, monitoring of internal activations, Gray Swan IPI lead on indirect prompt injection, chain-of-thought monitors that can stop a run, and sealed sandboxes before high-risk training. Whitwam’s hed: “So much for Gemini 3.5 Pro.” Google promised that model in June, he wrote; the summer was Flash. This file is Argon, not 3.5 Pro.


