Google Makes Gemini 4 Argon Generally Available, Two Weeks After Giving It Only to Cyber Defenders

Google made Gemini 4 Argon generally available on October 13, two weeks after restricting it to roughly 650 vetted cybersecurity organizations through its Fairwind Program with the model's own guardrails switched off -- full capability, aimed at defenders who needed to see what an unrestricted model could do against real attackers, before anyone else got to use it at all. The wider release is narrower than "generally available" usually implies: developers and enterprises reach it through the paid API in AI Studio and Vertex, consumer access depends on which subscription tier a user already pays for, and the headline price is temporary.

The staged rollout is itself a small data point about how a model with real offensive-security capability now reaches the public. Rather than one release date for everyone, Argon shipped in two tranches three weeks apart: a narrow, high-trust, guardrails-off group first, then a broad, guardrails-on public tier second. That ordering lets a company claim an early release date while actually stress-testing the riskiest capability on the smallest, most vetted audience first, before the same underlying model reaches everyone else with its guardrails back on.

Argon launches at $2 per million input tokens and $10 per million output tokens, with cached input priced 95% lower. That rate holds only through November 30, after which it rises to $4/$20. Google AI Ultra subscribers get Argon in the Gemini app, in AI Mode in Search, and in Antigravity, Google's coding environment; Google AI Pro subscribers get it only in the Gemini app and Search, not the API or Antigravity -- a distinction easy to miss in a headline that just says "available."

Argon's strongest independent result is real: on the Vals Index, a benchmark run by Vals AI -- an evaluator with no commercial stake in any lab's results -- Argon ranks #1 overall at 68.90%, just ahead of Claude Sonnet 5.5 (67.04%) and Claude Opus 5.5 (66.97%). The index weights performance across finance, coding, legal and tax tasks by each sector's share of the US economy, so it's measuring economically useful work on specific job categories, not general reasoning. Google also points to an 87% score on embodied reasoning (ERQA, its own selected benchmark), and a Gray Swan AI red-team test that found just a 0.7% success rate for indirect prompt-injection attacks across 15 adaptive attempts.

The Artificial Analysis Intelligence Index, a broader independent measure of general-purpose capability, tells a different story at the top: Argon ties GPT-6 Astra and Claude Fable 5.1 at 53, the same score it posted in its narrower Fairwind-only preview three weeks earlier, and sits five points behind Claude Opus 5.5's 58. Both rankings are real and independently produced; they aren't actually in conflict, because they're measuring different things. Vals Index rewards strength on a specific basket of economically weighted tasks, where Argon's pricing and specialization apparently pay off; Artificial Analysis aggregates a wider span of reasoning benchmarks, where Opus 5.5 still leads. A reader who sees only Google's "#1" headline misses that the honest claim is narrower than it sounds.

“The honest claim is narrower than it sounds.”

Not everyone inside Google is convinced the gap is just a matter of which benchmark you pick. Bloomberg reported that employees with early access to Argon found it strong on standardized tests but weaker on practical coding and front-end design work -- the kind of messy, real-world software task a benchmark suite doesn't fully capture, a pattern critics call benchmaxxing. Edwin Chen, founder of the AI-evaluation firm Surge AI, compared it to a student who earns a top SAT score without developing the skills the test is supposed to predict. Koray Kavukcuoglu, who runs day-to-day operations at Google DeepMind, rejected the comparison and said he remains confident the company stays at the frontier.

Argon also replaces a model Google never shipped. Bloomberg Intelligence analysts estimate the company may have written off up to $400 million in training costs by scrapping Gemini 3.5 Pro in favor of Argon -- a figure Google has not confirmed, and one that would only show up, if at all, in a future earnings disclosure rather than a launch blog post. Internally, supporters point to Argon's strength in safety, cybersecurity defense and natural conversation, and to Google's consumer chatbot and AI Mode in Search having together passed 1 billion users -- a distribution advantage no benchmark score captures, independent or otherwise.

The Fairwind cyber-defenders who got Argon first, guardrails off, three weeks ago keep that arrangement; this release adds a second, more restricted tier on top of it rather than replacing it. For everyone else, the model that shipped this week is priced to be cheap for seven weeks, genuinely leads one real independent benchmark and is tied rather than ahead on the broader one, and -- per Google's own employees -- may not yet close the gap between what it scores and what it ships for a working engineer. All three things are true simultaneously, and none of them is the full picture on its own.