Gemini 4 Argon is out: Google is back on top, but not everywhere

Google unveiled Gemini 4 Argon on 30 September as its new frontier model. Since 13 October it is actually available: in the Gemini app and Search’s AI Mode for Google AI Pro and Ultra subscribers, in the API for paying developers, and in Antigravity, Google’s coding environment, for Ultra subscribers. Free accounts don’t get it yet.

What it can do

According to Google’s own figures, Argon leads some “agentic” coding tests, where the model has to solve real problems in a software project on its own, ahead of Claude Opus 5.5 and GPT-6 Astra. On Arena’s text leaderboard, where people vote, it has moved into first place, although the result is still marked preliminary.

The picture isn’t uniform, though. On other coding tests it trails its rivals, and on the Artificial Analysis intelligence index, which combines many independent benchmarks, it doesn’t beat Claude Opus 5.5.

What it costs

Via API the launch price is $2 per million input tokens and $10 per million output until 30 November; from 1 December it doubles to $4 and $20. Repeated prompt parts read from the cache cost much less. For everyday users another figure matters: you need at least Google AI Pro, €21.99 a month.

My take

It’s an excellent model and, for people who live in Gmail, Drive and Docs, the news is concrete: Google’s most capable model now works inside the tools they already use. I wouldn’t call it the new undisputed king, though: for writing code and working on complex documents, Claude is still, in my view, the most solid choice, and the top spots change every few weeks.

The practical advice is always the same: if there’s a task you do often, try it on Argon and on your current model with the same examples, and decide on results, not headlines. I explain how in the lesson on evaluating AI with your own tests.