October 16, 2026
Google made Gemini 4 Argon broadly available on October 13, 2026, following a cybersecurity-only preview that started September 30. As VentureBeat reported, developers can reach it through the paid Gemini API, Google AI Studio, Android Studio, Firebase AI Logic, and eligible Gemini Enterprise plans. Introductory API pricing runs $2 per 1 million input tokens and $10 per 1 million output tokens through November 30, 2026, then jumps to $4 and $20 after that. Cached input gets a steep 95% discount, landing around $0.10 per 1 million tokens. MarkTechPost independently confirmed the release date and pricing structure in its own coverage.
Here's the detail easy to miss in the pricing table: the two capabilities most relevant to actually building something with Argon -- Google AI Studio's "Build" agent flow and Antigravity, Google's agent-first dev platform -- are not available to Google AI Pro subscribers. Per VentureBeat's reporting, individual access to Antigravity requires the $99/month Google AI Ultra tier. Pro subscribers get Argon inside the regular Gemini app and AI Mode in Search, which is a materially different product. So the real stack to go from idea to working agentic build, as an individual developer, is a $99/month subscription, plus metered per-token API costs on top if you touch the API directly for anything beyond the subscription surface.
Google is reporting strong internal adoption: more than 100,000 employees reportedly use Argon through an internal Antigravity build monthly, with a 35% rise in agentic code submissions and a 30% drop in post-submission rollbacks, according to VentureBeat's coverage of Google's self-reported figures. Those are Google's own numbers about Google's own employees using an internal build -- useful as a signal of intent, not as independent validation.
Outside reception of the tooling is more measured. AI researcher Sam Witteveen of Red Dragon AI said on record, "I still don't think, for example, Antigravity is anywhere near as good as Claude Code or Codex," while separately calling the model itself strong on generalization. That's an important distinction: the critique is aimed at Google's agentic tooling layer, not Argon's underlying capability. On Terminal-Bench 4.0, a command-line coding benchmark, Google's own comparison numbers show Claude Opus 5.5 beating Argon 66.4% to 57.4%. MarkTechPost adds that independent benchmark reception is mixed more broadly -- Vals AI ranks Argon first on its index, while Artificial Analysis gives it an Intelligence Index of 53, below Claude Opus 5.5.
Step back from the benchmark scores for a second. The pattern across Google's launch is one worth naming plainly: the biggest labs are turning "AI builds your app or agent for you" into a premium tier stacked on top of metered API usage, rather than something a team can just pick up and use. You need Ultra to unlock the agentic building tools at all. Then, even with Ultra, you're paying per-token for anything that touches the API, at a price that's scheduled to double after November 30. And after all that stacking, independent reviewers are telling you the tooling layer itself -- the part that actually matters when you're trying to ship something -- still lags dedicated coding-agent products like Claude Code and Codex.
For a business trying to decide whether to build its own internal tool or customer-facing app, "which foundation model wins which benchmark" is close to the least useful question you could ask. The question that actually determines whether a project gets finished is: what does it cost, in subscription tiers, API metering, and tooling maturity, to get from an idea to something that works and keeps working? A model that scores well on Terminal-Bench doesn't help if the agent harness built on top of it is the weak link, and a cheap per-token rate doesn't mean much if you can't touch the agentic features without first paying for the top consumer tier. We've made a version of this argument before when comparing per-seat SaaS costs against building your own -- the sticker price is rarely the real price.
This is also the structural reason ViibeStack isn't a bare model subscription you have to wire into your own agent harness. When you're evaluating options, it's worth separating two very different categories: a foundation model API plus a coding-agent tool you assemble yourself, versus a dedicated app-building platform where getting started doesn't require layering a premium subscription tier on top of token metering just to reach the parts that actually build things. ViibeStack is built around predictable scope from the start -- you're not guessing whether a feature lives behind a higher tier or whether your bill doubles next month because an introductory rate expired.
None of this means Argon is a bad model -- even its critics, including Witteveen, credit it with strong generalization. But a strong model wrapped in an immature, gated tooling layer is still an immature, gated tooling layer. If you're deciding how to actually build the software your business needs, it's worth reading our buy vs. build vs. ViibeStack breakdown and thinking about total cost and time-to-working-product, not which lab's model wins this month's benchmark.
Sources