Google made Gemini 4 Argon generally available on October 13. I covered the model’s benchmarks at the start of October, when it was an announcement without availability. Now it’s live worldwide through the API in AI Studio and on the Gemini platform, for Ultra subscribers in Antigravity, and for Pro and Ultra subscribers in the Gemini app and AI Mode in Search. Whether Plus and free accounts ever get it, Google won’t say.
The price is the actual news
Argon launches at $2 per million input tokens and $10 per million output tokens, with cached input 95 percent off. By The Verge’s reckoning that’s roughly a fifth of what OpenAI’s GPT-6 Astra and Anthropic’s Fable 5 cost. The introductory period ends in late November, when those rates double. Even after that, Argon stays substantially cheaper per token than its frontier rivals, which doesn’t automatically mean cheaper bills: two models can burn very different numbers of tokens on the same job.
For rankings Google points at the Vals Index, which scores models across finance, coding, legal and tax work. Argon sits at number one there. Google also claims Argon beats GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across a range of benchmarks covering coding, cybersecurity, science, finance and law.
What Google did with it internally
The in-house examples are more interesting than the benchmarks. With Argon in Antigravity, Google says teams increased agentic code submissions by more than 35 percent while cutting code rollbacks by nearly a third.
Inside Chromium’s 40-million-line codebase, the Chrome team used Argon to separate a legacy folder of roughly six million lines from the core page-handling code, across hundreds of files. Google describes that as months of work finished in weeks. The payoff for users is meant to be live pages moving between tabs, side panels and floating windows without reloading.
Other numbers Google puts forward: in the Agent Development Kit, in-memory session prep in long sessions dropped 40 percent and cold imports 21 percent. The quantum chemistry package xTB got 2.3x faster at predicting molecular structures. And on Gemma 4 E2B, Argon pushed throughput via vLLM from about 37,900 to over 108,000 tokens per second, with no quality drop by Google’s account.
For Claude users, the date to watch is November
The benchmark claims come from Google and will meet outside measurement over the next few weeks. The price is the solid part, and it aims straight at where agents get expensive: many tokens for little visible output. Anthropic moved on exactly that recently, cutting prompt cache reads on Sonnet 5.5 from $0.20 to $0.10 per million tokens in early October. Late November, when Google’s introductory rate expires, will show how much of the gap survives.