Gemini 4 Argon is with a short list of cybersecurity defenders. It is not in the public Gemini app.

CNBC quotes Tulsee Doshi, who leads Gemini model product. A limited start "gives us more confidence" and gets "a model that is trained and strong in cyber defense in the hands of defenders as soon as possible." CNET names the list: Google's Fairwind program. Asked when the rest of the world gets the model, Google did not answer.

Hundreds of terabytes is not a bill

Google says Argon is already optimizing memory in its own data centers, "freeing up hundreds of terabytes of memory without buying additional hardware." "Hundreds" means at least 200 only if you take the word at its smallest, and the company may mean far more. It did not publish terabytes as a precise count, and it did not publish a price. There is no dollar figure to divide into.

CNBC says Gemini 4 comes nearly a year after Gemini 3, the release that put Google back at the front of the model race. CNET says a lot changed in the 10 months since Gemini 3. Nearly a year and 10 months are two clocks for one gap. They point the same direction and they are not interchangeable. A rumored Gemini 3.5 Pro earlier this year did not ship. CNET says Google held it back to keep working. Bloomberg, via CNET, reported worries inside the company this week that Gemini 4 was still weak at coding.

Four safeguard areas stand between this and a public model

Before a public launch, Google said on Wednesday it wants safeguards scaled in four cybersecurity areas, misuse and prompt injection among them. Doshi called Argon "incredibly well-rounded" and good at long, multi-step work. CNBC says it beats the 3.8 Flash Cyber model from earlier this month at finding vulnerabilities. On cyber benchmarks it ties OpenAI's GPT-6 Astra and Grok 4.7, and it leads GPT-6 Astra and recent Anthropic models on the Vals Index. CNET says Google's tables against Astra and Anthropic's Fable 5.1 support the claim of strength in software engineering, cybersecurity, and legal and finance work. The output limit CNET lists is 1 million tokens.

A Google blog post, quoted by CNET, says the company is "actively engaged" with the White House voluntary review for new models. OpenAI kept GPT-6 Astra off a general release. Anthropic did the same with Fable 5. OpenAI scrapping a model release over safety is the same posture. A Gemini system that got out and struck three companies is the reason a cyber model starts with defenders.

Fairwind is not a date

Argon codes, reads images and charts, and reasons. If past launches are a guide, CNET expects some Gemini 4 to land inside Search and Docs. That is a pattern, not a ship date. Google also says the model is in use on quantum research. The memory-optimization line and the "hundreds of terabytes" line are one internal job described two ways.

Wednesday's announcement has a user group and no calendar. Fairwind has the weights. The public date is the question CNET asked and Google left blank.

Doshi's "as soon as possible" applies to defenders, not to Search. The four safeguard areas are the gate Google named, and prompt injection is only one of them. CNET lists coding, images, charts, and reasoning as things Argon can do, and business tasks in engineering, security, law, and finance as where Google says it is particularly adept. Benchmarks against Astra, Grok 4.7, and Fable are comparisons Google and the two outlets are willing to print. A coding doubt inside the company, via Bloomberg, is the counter-note in the same week. Both can be true: a model that ties Astra on a cyber test and still worries its own engineers on ordinary code. The release that would resolve the tension is the one without a date.