Google launches Gemini 4 Argon

A staged release

Gemini 4 Argon did not go straight to the public. Google announced the model on 30 September 2026 to trusted cyber defenders taking part in its Fairwind Program, so that its first users were people whose job is to look for weaknesses in software rather than general customers^[1]. Releasing a model to a narrow group first is a common pattern: it lets the maker watch how the system behaves on real, consequential work while the audience is still small and the feedback loop is short. For a security partner, the interesting question is not whether a model can write fluent text but whether it can spot something a human or an earlier system walked past.

One result of that kind was made public at launch. Wiz reported using Argon to find a critical vulnerability in healthcare software that earlier frontier models had missed^[1]. A find like this is narrow evidence — one case, one vendor, one class of target — but it is also the kind of concrete comparison that early announcements rarely contain, because it sets the new model against predecessors on the same task.

Worldwide availability

On 13 October 2026 Google made the model available worldwide to developers, enterprises, and Google AI Pro and Ultra subscribers^[2][3]. Three quite different audiences are served at once: programmers building on the API, organisations deploying it internally, and individuals who reach it through a consumer subscription. Google described the release as the start of the Gemini 4 family and named Gemini 4 Carbon as the next frontier model to come^[2]. Naming a successor before the first member has settled is a statement about schedule as much as about capability: the generation is presented as a line of models rather than a single product, with the members expected to differ in size, cost and the tasks they are aimed at.

Introductory pricing

For developers the launch came with introductory API pricing of $2 per million input tokens and $10 per million output tokens^[2]. Tokens are the pieces of text a model reads and writes — words are split into fragments, and every fragment counts. Input is what you send to the model, output is what it writes back, and output is priced higher here because producing text is the part that costs the provider most compute. Listing the two numbers separately lets a team estimate a bill from its own traffic before committing, which matters more for high-volume applications than for experiments.

Access for researchers

Scientists were offered a separate route in: they could apply for six months of free access to Antigravity's highest usage tier, with 10,000 places available^[2]. Usage tiers normally cap how much of a service an account may consume, so granting the top tier for half a year removes the budget question for accepted applicants and lets them plan longer projects rather than short trials.

What the launch shows

Read together, the details describe a release in which the first audience was security, the second was paying developers and enterprises, and the third was researchers given quota rather than discounted rates. The Fairwind stage is the most distinctive part: by starting with defenders and publicising a vulnerability find, Google framed the model's advance in terms of what it can locate in existing code, a claim that can be checked against earlier models on the same target. The naming of Gemini 4 Carbon extends the same framing forward, presenting Argon as the opening of a generation rather than a finished story.