On 2026-10-13 Google's Gemini API models page gained a new entry under "Stable": gemini-4-argon, tagged "New Stable", with a banner that reads "Gemini 4 Argon is now available." The same day DeepMind put up its model page and Google's developer blog published "Gemini 4 Argon for Robotics". On Hacker News, "Gemini 4 Argon" passed 1021 points.
Until this week Argon was something Google talked about. Now it is a model ID you can put in a request. This post covers what Google shipped, what it costs before and after December 1, where Google's own benchmark table says it wins and loses, and what is still not written down anywhere.
What happened
Google describes Argon on the models page as "Our flagship reasoning model with a unified 1M context window and long decodes up to 1M tokens." There is one endpoint, gemini-4-argon. No preview suffix, no dated variant.
Access comes in three doors, and they do not line up neatly:
- The Gemini API. The pricing page says "Gemini 4 Argon is not available on the Free tier or on the Google AI Pro plan." It is on the paid tiers only: Standard, Batch, Flex and Priority.
- The apps. DeepMind's page says "Gemini 4 Argon is available today for Google AI Pro and Ultra subscribers", in the Gemini app, Google AI Studio and Google Antigravity, and that it is "expanding to hundreds of additional Fairwind Program partners today."
- The SDKs. The robotics post gives the model ID in the Interactions API and promises "day-one support across the google-genai SDKs", with AI Studio and Antigravity access named for Google AI Ultra subscribers.
So if you build on the API, you need a paid key. If you only want to try it in a chat window, a consumer plan may be enough, depending on which Google page you read.
How it works, as far as Google says
Google does not publish an architecture. What the pages describe is the shape of the thing: a reasoning model, one context window of 1M tokens shared between what you send and what it writes back, and output that can run very long. "Long decodes up to 1M tokens" is the unusual part. Google is saying Argon can, in principle, write as much as it reads.
The robotics post adds one more claim worth knowing: 87% on ERQA, a spatial and physical reasoning test, "using pure reasoning". That is a robotics number, but it says something about the model's general direction: Google is selling Argon as a thinker first.
One detail on the pricing page matters more to a web app than any benchmark: "Implicit caching only (explicit caching is not supported)." If your app keeps a long system prompt warm with an explicit cache handle, that pattern does not carry over to Argon. You get whatever the platform caches on its own.
What it costs, and when that changes
This is the part to write down before you estimate anything. The pricing page prints a clock.
| Tier | Input / output per million tokens, through November 30, 2026 | From December 1, 2026 |
|---|---|---|
| Standard | $2 / $10 | $4 / $20 |
| Batch and Flex | $1 / $5 | $2 / $10 |
Cached input on Standard is $0.10 per million tokens through November 30. Output pricing includes thinking tokens, so a model that reasons at length bills for that reasoning too.
The practical meaning: every Standard and Batch price doubles on December 1. A cost estimate you make this month is wrong in about seven weeks. If you are pitching a feature to a client or a manager with Argon inside it, quote the December price, not today's.
Where Google's own table says it wins and loses
Every number in this section is Google's, from DeepMind's page. No independent evaluation was on the pages I read.
Where Argon comes first:
- Vibe Code Bench: 91.9%, ahead of Opus 5.5 and Fable 5.1 (both 90.3%) and GPT-6 Astra (89.6%).
- DeepSWE v1.1: 77.9%, against Opus 5.5's 74.2%.
- AutomationBench: 51.3%, against Opus 5.5's 42.5%.
- GraphWalks 256k to 1M BFS (F1), a long-context test: 84.2%, against Opus 5.5's 66.8%.
Where it trails:
- FrontierSWE v2: 55.0%, behind GPT-6 Astra (65.5%), Opus 5.5 (62.3%) and Fable 5.1 (56.3%).
- Terminal-bench 4.0: 57.4%, behind Opus 5.5's 66.4%.
Read as a whole, the table is mixed for the work coding agents do. Argon leads on the benchmarks closest to "build this small app" and on long context. It trails on the ones closest to "work through a large repo in a terminal". A vendor's own table usually flatters the vendor, and this one still shows two clear losses. That is worth taking seriously.
Why someone building web apps with AI should care
- It is real now. You can call it from the
google-genaiSDK on a paid key today. Before 2026-10-13, planning around it was guesswork. - The long-context result is the interesting one. If your product reads whole codebases, long transcripts or big document sets in one go, the GraphWalks gap is the number to test first.
- The price clock changes the maths. A model that is cheap to try in October may not be the cheap choice in December.
- Caching works differently. Without explicit caching, a chat product with a long fixed prompt may cost more on Argon than the headline price suggests.
- Test it on your repo, not on the chart. The benchmarks split by task type, and only your own tasks tell you which side of the split you are on.
What is not known yet
- Which plans get what. DeepMind's page says Google AI Pro and Ultra; the robotics post names Ultra for AI Studio and Antigravity; the pricing page's "not on the Google AI Pro plan" is about the API. No single Google page reconciles them.
- Rate limits for the paid tiers. Not on the pages I read.
- Vertex AI availability. DeepMind's page does not say.
- Independent evaluations. Every benchmark above is Google's own.
My take
I build frontend apps with AI agents every day, and what I care about is the boring part: does the agent finish the task in a real repo, and what does a month of it cost. On both, Argon's own launch material tells me to wait a little. The terminal and large-repo numbers are where my work lives, and those are exactly the rows where Google shows it behind. The long-context result is the one I want to try, because reading a whole Next.js app at once is where today's tools still lose the thread.
What I like is that Google printed the price change in advance. Most launches do not. It means I can plan with the December number from day one instead of finding out from an invoice. My rule for a new model is the same as always: one real task from my own backlog, the same prompt on two models, and the bill read back afterwards. Argon gets that test before it gets anything else.
