Gemini 4 Argon nears public release: it reads far more, but big inputs cost up to 1.8x
Leaked code suggests Google will open Gemini 4 Argon to everyone, with pricier tiers for longer documents. Learn what that means for your costs and why you shouldn't plan around it yet
Today in AI (2026-10-12)
- Google's Gemini 4 Argon shows signs of a public release: Google's restricted top model may soon be usable by any startup, with options to read much larger documents at higher cost.
- Amazon in talks to buy AI chip-switching startup Decart for ~$7B: Big cloud providers are paying billions to make AI cheaper to run, which could lower the costs of AI features in your product.
- Anthropic opens Claude Code Projects to all Pro and Max users: Founders on a paid Claude plan can now have Claude run several coding tasks in parallel inside one ongoing project.
Short answer: probably yes, Gemini 4 Argon is heading toward a public release, but nobody outside Google knows when. Don't plan your product around it until it ships and you've tested it on your own work.
Here is what happened. Code changes in Google's Antigravity coding tool now show Gemini 4 Argon in the model menu, according to TestingCatalog. That suggests Google is preparing to release a model it had restricted to security teams. Google has not announced a date or confirmed pricing.
What Gemini 4 Argon is
Gemini is Google's family of AI models, the same kind of technology behind ChatGPT and Claude. Argon is the most powerful one Google has built, and until now only security teams could use it. I covered why it was locked down in this earlier post.
The leak is a hint that the lock may come off. Think of it like seeing a new menu item appear on a restaurant's internal ordering screen before the restaurant announces it. Strong sign. Not a promise.
What the leak does and doesn't confirm
It suggests:
- Argon is being wired into a tool normal developers use.
- Google is building a pricing structure for it.
It does not confirm:
- A release date.
- Final prices.
- That the version you can use will match the one security teams have.
- That it will be better at your task than the model you use today.
Working memory comes in three sizes, and bigger costs more
Every AI model has a "context window," which is how much text it can hold in mind at once: your documents, your instructions, the conversation so far. Think of it as the size of the desk the model works on. A bigger desk lets it read a whole contract set or a whole codebase in one go instead of in pieces.
The leaked setup has three desk sizes:
| Working memory | Cost per request |
|---|---|
| 256K (default) | Baseline |
| 512K | About 1.3x |
| 900K | About 1.8x |
"K" means thousands of tokens, the small chunks of text the model reads. As a rough guide, you can think of it as the length of the material it can take in. The 900K option holds roughly 3.5 times the default for 1.8 times the cost. So per unit of text read, the big option is cheaper. But if your users send it large inputs on every request, your bill still climbs by up to 1.8x.
One caveat: these multipliers refer to quota use per turn inside Antigravity, a coding tool. How Google prices this for products you build on top of it may differ. Treat the numbers as a signal about the shape of the pricing, not a quote.
Thinking effort is a second cost dial
Google's Gemini web app also added low, medium and high "thinking effort" settings. This controls how long the model reasons before answering. Higher effort usually means better answers on hard problems, and slower, more expensive ones.
So you will likely have two dials: how much the model reads, and how hard it thinks. Most products don't need both turned up. A support chatbot answering simple questions should run low. A tool that analyzes a 300-page diligence file might justify the largest window and high effort.
What this means for you
If you're building something now, don't wait for Argon. I made the same point when OpenAI shelved GPT-6.1 Astra: a model that isn't released can be delayed, changed, or cancelled. A roadmap that depends on it is a roadmap that depends on someone else's schedule.
What I would do instead is build on a model you can use today, in a way that lets you swap later. Here is the checklist to run with your developer or agency:
- Can we swap models in a day or two? The AI provider should sit behind one small piece of code, not be woven through the product. Good answer: "Yes, it's one configuration change plus a test run." Bad answer: "We'd have to rework a lot."
- Do you have a set of test cases from our real work? Ask for 20-50 real examples with known good answers. When Argon ships, you run them against it and compare quality and cost with evidence instead of marketing.
- How do we control cost per request? Ask whether you can cap how much text goes into each request and set the thinking effort per feature, not globally.
- Do we need the big window at all? Ask what the largest input your users actually send is. If it fits in the default size, you don't need to pay more.
- What's our backup if one provider raises prices or pauses a launch? There should be a named second option, even if it's not built yet.
Common mistakes
- Promising customers a feature that needs an unreleased model. If it slips, you slip.
- Choosing the biggest window "just in case." You pay for it on every request.
- Judging by benchmarks. Public rankings tell you little about your documents and your users.
- Switching on day one of release. New models often have rough edges in the first weeks. Test first.
What to do this week
Send your developer the five questions above. Then ask for one deliverable: a small set of real test cases and a note on how long a model swap would take. If the answer is "weeks," fixing that is worth more than any single model release. And if you're weighing another recent launch, my post on GPT-6 applies the same test.
Bottom line
Argon looks close, but a menu entry is not a launch. Build on what you can use today, keep the ability to switch, and let your own test cases decide when Argon earns a place in your product.