Gemini 4 Argon in Antigravity: Better Instruction Following on Google AI Ultra

I tried Gemini 4 Argon both in the Gemini app and in the Antigravity app. Unlike earlier models, it followed instructions more accurately, ran the tests that were needed and delivered the work more correctly. On my Google AI Ultra 5x plan, the Argon quota has felt quite generous.

These are impressions from my own use, not a measured comparison, and I did not record the dates, tasks or thinking level. The quota remark describes how my plan has felt in my usage, not a documented allowance.

What the official guidance supports

Google's Antigravity blog post of October 13, 2026 says Gemini 4 Argon is rolling out inside Antigravity for Google AI Pro and Ultra subscribers, in the Antigravity desktop app and the Antigravity CLI. For businesses, it is offered through the API in the Gemini platform and inside Antigravity on eligible Gemini Enterprise subscriptions. On October 15, the official Antigravity account on X said the rollout to paid Pro subscribers was happening that day.

The Antigravity models page lists Argon for Google AI Pro, Ultra and Enterprise, but not for the free tier or Google AI Plus. On Pro it is marked as available for non-trial subscriptions only. It is also marked as unavailable in the Antigravity IDE; the page names the Antigravity App, the CLI and IDE extensions instead.

What Google reports from its own use

The blog post says more than 100,000 users across Alphabet use Argon in an internal version of Antigravity each month, and that over four weeks Google saw a 35% increase in agentic code and a 30% decrease in post-submit rollbacks. It also describes internal cases such as closing 97 of 134 stale or duplicate bug reports in under three hours. These are Google's own figures, not independent measurements.

How it sits next to Gemini 3.8 Flash

The earlier Field Note on Antigravity covered Gemini 3.8 Flash and Remote Control. The plans page still lists Gemini 3.8 Flash among the baseline features for every plan. Pro quota refreshes every five hours until a weekly limit is reached, and usage beyond the baseline can be paid with AI credits when the overage setting allows it. Argon adds a heavier option on top of that rather than replacing the everyday model.

I did not compare Argon directly with Gemini 3.8 Flash, so this note does not say on which tasks Flash remains enough. My comparison is with earlier models in general.

Where I would draw the line

My current conclusion is that Argon is a noticeable improvement over earlier models in how accurately it follows instructions, tests its work and delivers it, at least in the work I gave it. On Ultra 5x the Argon quota has felt quite generous to me, but I cannot say how the same work would feel on Google AI Pro, which the plans page places below Ultra.

A useful next test would record the plan, the Antigravity surface used, the thinking level, the task, and how much of the plan's quota a typical Argon session consumed compared with Gemini 3.8 Flash.