Google appears to be preparing a wider release of Gemini 4 Argon, the model it gave only to vetted cybersecurity teams on 30 September. TestingCatalog reported on 9 October that recent changes to Antigravity add Argon to the product’s model selector, along with settings that show how much of a user’s allowance each conversation will consume. Google has said nothing about any of it.
On 1 October we reported that Argon was limited to trusted cyber defenders and that Google offered no date for anyone else. This is the first outside evidence that the gap may close. It comes from a leak in shipped code, not an announcement.
The numbers are the useful part. TestingCatalog found three context configurations. The default is 256K. A 512K option costs about 1.3 times the usual quota on each turn, and a 900K option costs around 1.8 times. Both multipliers are approximate in the report.
A context window is the amount of text a model can hold in view at once: your instructions, the files it has read, and everything said so far. A turn is one exchange, your message and the model’s reply. Quota is the allowance a subscription gives you, and each turn draws it down. Using the 900K setting would burn nearly twice the allowance per exchange. If a plan covers 100 turns at the default, the same plan would cover roughly 55 at the largest size, assuming the multiplier holds. That arithmetic is ours, not the source’s.
TestingCatalog adds that these options resemble existing choices for Gemini Flash models, and that how context size relates to reasoning effort is unclear. It also does not explain what Antigravity is. Its reporting links the product to Google’s work on autonomous coding agents.
Separately, Gemini Web now lets users set thinking effort to high, medium, or low on the models it already supports. TestingCatalog suggests this could prepare the app for Gemini 4. No evidence ties the two changes together.
A selector entry shows that engineers are wiring a model into a product. It does not show a decision to ship, a price, or a date. Google has given no date, and no outside report changes that.
The shift in audience is the larger point. Google first pitched Argon as a tool for vetted defenders, partly because it is good at finding software flaws, and a model with that profile moving toward ordinary subscribers means the cost of long sessions becomes an everyday budgeting question. Until Google publishes terms, treat the multipliers as a preview of the pricing logic, not a price list.
Teams that want Argon for large codebases should budget at the 900K multiplier, not the default, because long-context coding work is the likeliest reason to pick that setting.
Reported by TestingCatalog (Alexey Shabanov), 9 October 2026.
