Gemini 4 Argon arrives 333 days after the last flagship, and you cannot use it

Alphabet released Gemini 4 Argon on Wednesday and called it its most powerful model. Access runs through the Fairwind Program to a selected group of cyber partners, so for everyone else the release is a blog post.

The calendar, closed out

We counted the gap a week ago, when Koray Kavukcuoglu said Gemini 4 was in refinement and would land much earlier than the end of the year. The counters now read:

• 333 days since Google's last flagship, the Gemini 3 series of November 2025
• 92 days left in 2026 when Argon arrived, which makes "much earlier than the end of the year" technically correct
• 71 days since Google shipped the same capability once already, on 21 July
• Two gated cyber models in one quarter, neither available to the public
• One billion monthly users on the Gemini app, level with ChatGPT

Look at the 21 July release. Gemini 3.5 Flash Cyber was fine-tuned to find, validate and automatically patch vulnerabilities in code, and Google restricted it to governments. Argon is described as able to autonomously find, validate and patch critical software vulnerabilities, and Google has restricted it to Fairwind partners. The sentence is nearly word for word the same one, seventy-one days apart, attached to a model a tier higher.

Flash Cyber came with a number behind the claim: inside CodeMender it found 55 unique confirmed vulnerabilities while being tested on Chrome's V8 engine. Argon arrives without an equivalent figure.

Benchmarks, announced by the vendor

Google says Argon scored higher than OpenAI's GPT-6 Astra and Anthropic's Fable and Opus across a range of benchmarks, and cites the startup Vals to show Argon leading its model index.

“Built to sustain deep reasoning across complex, long-horizon workflows, Argon is fundamentally changing the way we work and build at Google.”

— Google, Запись в блоге, 30.09.2026

Google, in its blog post on Wednesday

Hold that against what the same company proposed in August, when Google DeepMind piloted cryptographic double-blind testing so that an evaluator cannot tell which lab's model it is scoring. Six weeks later the leadership claim arrives in a company blog post citing a third-party index. Both can be sincere. Only the double-blind test produces a number someone outside Google can repeat.

What Argon actually does

Beyond security, Google puts Argon on coding and engineering work and says its own staff already use it for debugging and codebase migrations. Multimodal AI is part of the pitch too: the model parses long videos and charts. Google sells it as sustained reasoning over long workflows.

Distribution is the new part. OpenAI held back a flagship after testing found it acting without permission. Google is shipping its flagship to a security programme and nobody else. Two labs, two reasons, and the most capable model announced this quarter is one you cannot open.

Scale sits oddly beside that. Google said in August that the Gemini app passes a billion users a month, which puts it level with ChatGPT. A billion people use the models Google does release; the one it calls most powerful goes to a list.

Informational material, not investment advice. Benchmark claims come from Google's own announcement and have not been independently reproduced.