Gemini 4 Argon launches globally — but stumbles on agentic coding in Google's own table
Google's largest model release in nearly a year was first opened to a small group of cybersecurity practitioners. From October 13, Gemini 4 Argon is rolling out worldwide — with an unusually long output limit, an introductory price that makes it one of the cheapest in the frontier class, and a benchmark portfolio that Google itself shows is mixed.
What has happened
Google first announced Gemini 4 Argon on September 30, 2026, in a blog post by Google DeepMind's SVP Koray Kavukcuoglu. At that point, the model was limited to trusted cybersecurity defenders in Google's Fairwind Program (Mashable). The global rollout came on October 13.
According to The Verge, Argon is Google's first major frontier release since Gemini 3 in late 2025 (The Verge) — that is, the first substantial answer to competitors in more than nine months.
What developers actually get: one million tokens out
The most important change for developers is not the input context, but the output length. Argon can write up to one million tokens in a single response — compared with 64,000 tokens on previous Gemini models (MarkTechPost).
The mechanics are worth understanding: the model card describes a combined window of up to one million tokens for input plus output together. A long prompt thus eats into the space left for output. For very long generations, the model card documents a continuation protocol that lets responses continue across multiple calls.
Otherwise, Argon is according to the model card an entirely new model, not a fine-tune of a previous Gemini: a large-scale sparse mixture-of-experts transformer with multimodal input (text, images, audio and video) and text-only output. The parameter count is not disclosed (MarkTechPost).
The price — and the ticking clock
The introductory API price is $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million (KuCoin/Shenchao TechFlow). By comparison, OpenAI's GPT-6 Astra charges $10 per million input tokens and $50 per million output tokens — that is, five times as much on output. With the introductory price, a full one-million-token response therefore costs $10 in output tokens.
But the discount has an expiry date. According to Ars Technica, the offer applies only through November 30, before the price doubles to $4 and $20 per million tokens (Ars Technica). The media are not entirely in agreement here, however: The Verge supports November 30, while Gizmodo writes that prices double "starting in December," and KuCoin states the promotional price applies "through the fourth quarter of 2026." The discrepancies cannot be reconciled from the available information — but the safe message for developers is that $2/$10 is temporary, and that the standard price afterward will be $4/$20.
Who gets access — and on what terms
For consumers, Argon is reserved for paying users: in the Gemini app, AI Pro or AI Ultra is required, while Antigravity and AI Studio offer the model only to AI Ultra subscribers. Google would not say when or whether Argon will come to lower tiers in Antigravity (Ars Technica).
The price of the Ultra subscription itself is also not consistently reported: Mashable states that Ultra starts at $100 per month, while Ars Technica writes $99. The detail is minor, but it shows that the figures in the coverage come from secondary sources, not from Google's own documents.
The benchmark picture: weaker on agentic coding
The honest picture is that Argon's strengths and weaknesses lie in different places — and Google itself lays this out openly.
In Google's own comparison table, Argon trails the competition on five rows. The most striking is FrontierSWE v2, where Argon scores 55.0 percent against 65.5 percent for GPT-6 Astra (MarkTechPost). Note at the same time that most rival scores in the table are self-reported or drawn from public leaderboards, and that Argon's DeepSWE score was self-computed — the numbers are thus not independently verified.
The most important third-party check is the Artificial Analysis Intelligence Index. There, Argon sits at 53 points with "high reasoning" enabled — on par with GPT-6 Astra, one point above GPT-6.1 Sol and Muse Wonder Max, but around five points behind Claude Opus 5.5 and Claude Sonnet 5.5 (Mashable).
That fits with the reporting around the launch: Gizmodo describes the debut as dampened by reports of skepticism among some Google employees about the model's performance in key areas such as coding, following earlier delays of Gemini 3.5 Pro (Gizmodo).
Google's own claims about internal use
Kavukcuoglu says Google itself uses the model internally: "At Google, we have used Argon on our hardest engineering problems," he writes in the launch post, pointing among other things to the model being used to streamline Chrome (Gizmodo).
The company also promotes concrete figures from internal use in Antigravity: reportedly over 35 percent more agentic code submissions and nearly a third fewer code rollbacks (Android Authority). These are company claims without independent verification — and they stand in some tension with the benchmark weakness on agentic coding that Google's own table shows.
The safety disclosure
Worth noting for a model that was first opened to cybersecurity defenders: Under Google's Frontier Safety Framework, the company states that it cannot rule out that Argon reaches the critical capability level in biology and cybersecurity. Google simultaneously reports that it has implemented safeguards it assesses as risk-reducing (MarkTechPost).
What remains unresolved
Three things developers should watch:
- API reach. Mashable points out that the launch announcement does not say whether API access has been expanded beyond Google's previous rollout plans.
- The pricing expiry. The conflict between November 30 (Ars Technica, The Verge), "starting in December" (Gizmodo) and "through Q4" (KuCoin) is not settled. Check Google's own pricing pages before planning around $2/$10.
- The primary sources. None of the figures in this story are drawn directly from Google's model card, blog post or developer documentation — everything comes via secondary coverage. The figures should be verified against the documentation before being used as the basis for production decisions.
The clear impression after the launch: Argon is a significant capability upgrade with an economically aggressive entry price and an output limit few others can match — but on agentic coding, the field many users care most about right now, there is still room for doubt, and Google itself appears to be among those who have it.