Gemini 4 Argon is out, and the price doubles Dec 1

Gemini 4 Argon is generally available as of October 13. If you want to try Google's newest frontier model at the launch price, the clock is already running: $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20.

The rollout, announced by Google DeepMind SVP Koray Kavukcuoglu, widens access well beyond the September 30 launch, which was limited to cybersecurity defenders through the Fairwind Program. That first announcement was covered here when there was no GA date and no pricing end date. Now there are both.

Where you can actually use it matters more than the headline. On the API side, Argon is on the paid tier only: the model ID is gemini-4-argon in AI Studio and the Gemini platform. Google's pricing page excludes it from the free tier and from the API side of the Google AI Pro plan, so do not assume your existing plan covers it. In the consumer products, it is rolling out to Google AI Ultra subscribers in Antigravity (desktop and CLI), and to Pro and Ultra subscribers in the Gemini app and AI Mode in Search. One caveat from the fine print: Google's subscription page limits advertised availability to selected countries, so the announcement does not guarantee it has reached your account yet. Check before you change plans.

The pricing structure rewards a specific behavior. Cached input tokens are 95 percent off during the intro period, ten cents per million, which makes Argon unusually cheap for agentic workloads that re-read the same context across many tool calls. That is presumably the point. Argon is positioned for "deep work": long-horizon reasoning, multi-step coding, finance, legal, and science, with what Google describes as a unified million-token input-output budget.

Now the obligatory skepticism, because the announcement is thick with Google's own numbers. The company claims Argon ranks first on the Vals Index across finance, coding, legal, and tax work. It claims internal teams using Argon in Antigravity increased agentic code submissions by more than 35 percent while cutting rollbacks by nearly a third. It cites a Chrome refactor completed in weeks instead of months, a 2.85x inference speedup on Gemma, and an 87 percent score on a robotics reasoning benchmark. These are Google's figures from Google's deployments. No independent test has been published. Treat the benchmark story as a reason to run your own evaluation, not as the evaluation.

The practical move is simple. If you run agent workloads and you have been waiting for a model ID to test against, the wait is over. Run your hardest multi-step tasks on Argon now, while the intro rate holds, and keep your current Pro-tier setup until your own numbers give you a reason to switch. The Fairwind Program, meanwhile, is expanding to hundreds more cybersecurity partners, which suggests Google still sees defenders as the showcase customer. Everyone else gets seven weeks of half price to decide.

What remains unknown: whether Google extends the intro pricing past November 30, what the rate limits look like (not published), and how the model actually performs on third-party agent benchmarks. Those answers arrive with independent testing, which has not happened yet.

Related: our September 30 piece on the Fairwind-limited launch, and Meta's Muse Wonder pricing split from yesterday for the other side of the frontier-model price war.