Google released Gemini 4 Argon on October 13, 2026, priced at $2 per million input tokens and $10 per million output tokens until November 30. On December 1 both rates double, to $4 and $20. That leaves anyone weighing Gemini 4 Argon for client work about seven weeks to try it at the lower rate. I want to use that window carefully.
In this article
- Why access, not capability, is the first thing to check
- What the Gemini 4 Argon price change does to a real brief
- Which agentic coding claims are Google’s own numbers
- Why the safety notes matter more than the leaderboard
- Five Gemini 4 Argon checks I would run before client work
Argon Arrives Behind a Paid Gate, Not a Free One
Argon is a paid-only launch, and that shapes who can test it this month. Google’s announcement says the model is “globally available today for developers and enterprises via the API in AI Studio and Gemini platform.” Consumers get it through Google AI Ultra and Pro subscriptions in the Gemini app and in AI Mode in Search.
The developer side is narrower than that headline suggests. The Gemini API pricing page lists the model as gemini-4-argon and states it is “not available on the Free tier or on the Google AI Pro plan.” Inside Google Antigravity, access is listed for Ultra subscribers.
Google is also running a separate track for researchers. According to the announcement, 10,000 scientists can apply for the highest Antigravity usage tier free for six months, subject to eligibility, capacity and selection. That is a generous offer, but it is not built for a consultant who wants to draft a client brief with Gemini 4 Argon next week.
That matters to me directly. I use Gemini every week for quick queries, summaries and cross-checks, on a mix of free and paid access. I also run Antigravity alongside Claude Code as a beta tester. Neither of those setups puts Argon in front of me today, so everything below is a read of what Google published, not a test.
The Gemini 4 Argon Price Doubles on December 1
Gemini 4 Argon is cheap for seven weeks and mid-priced after that. Here is the full table from Google’s pricing page, per million tokens:
| Item | Through Nov 30, 2026 | From Dec 1, 2026 |
|---|---|---|
| Input | $2.00 | $4.00 |
| Output (incl. thinking tokens) | $10.00 | $20.00 |
| Context caching | $0.10 | $0.20 |
Two details on that page are easy to miss. Output pricing explicitly includes thinking tokens, and caching is listed as implicit only. I covered why the first point matters in a post on hidden reasoning token spend: the reasoning you never see still lands on the bill.
To make the numbers concrete, take a research brief with 20,000 tokens of source material in and 3,000 tokens out. By my arithmetic, that costs about $0.07 at the introductory rate and about $0.14 from December. Neither figure is alarming for one brief. The real question is how much hidden thinking each call adds, and that only shows up once you meter it.
The window itself is the practical point. Any Gemini 4 Argon testing done before November 30 costs half of what the same test will cost in December. That makes October and November the cheap time to find out whether the model earns a place in a workflow, rather than the time to build a workflow around it.
Google also cites a 1M unified input-output token limit on the API. For long client documents, that is a useful ceiling, though a ceiling is not the same as reliable recall across the whole window.
Agentic Coding Claims Are Big, and Mostly Google’s Own
The headline coding results come from inside Google, so I read them as direction, not proof. The announcement says Argon ranked first on the Vals Index, which covers finance, coding, legal and tax work. That is an outside index, and it is the claim I weight most.
The rest are internal reports:
- In Antigravity, agentic code submissions rose more than 35%, and code rollbacks fell by nearly a third.
- A Chromium refactor that would have taken months finished in weeks, according to Google.
- Shortcut attempts on coding benchmarks fell to 2%, which Google calls its lowest for any Gemini model.
A model that ships more code is only useful if it also rolls back less, and Google put both numbers side by side for Gemini 4 Argon.
That pairing is the part I find persuasive. A rise in submissions alone could just mean more noise. A drop in rollbacks next to it suggests the extra output is holding up. When I looked at what Gemini 3.7 Flash meant for a coding stack, speed was the story. With Gemini 4 Argon, Google is selling durability, and that is a harder claim to fake for long.
Antigravity also picks up /goal slash commands and multi-agent loops with custom skills. Those are worth watching, but only for Ultra users right now.
The Safety Notes Matter More Than the Leaderboard
For client work, the most useful lines in the announcement are about behavior, not scores. Google says it trained Argon to “respect system boundaries, report negative findings honestly, and pause for human confirmation before taking irreversible steps.”
That is the behavior I want from any agent that touches a client folder. It also lines up with a second claim: on the Gray Swan indirect prompt injection benchmark, Google reports attack success cut to 0.7% across 15 adaptive attempts. Prompt injection is the risk I think about most when a model reads web pages or inbound documents on my behalf.
On writing, Google says Argon “follows style instructions better than any of our previous Gemini models.” That is a vendor claim with no number attached. It is also the easiest one for a solo operator to check, because you can hand the model your own style guide and see what comes back.
My standing rule still applies. I do not route client data to a new model on day one, whatever its safety card says. Gemini 4 Argon does not get an exception just because it comes from a vendor I already use.
Five Gemini 4 Argon Checks Before It Touches Client Work
The checks below are what I would run during the introductory window, on non-client material only:
- Confirm access first. Check whether your plan actually includes Gemini 4 Argon in the app, the API or Antigravity. The free tier and the Pro plan do not cover API use.
- Meter thinking tokens. Run the same three prompts at different lengths and record the real output token count, since thinking is billed as output.
- Test the style claim with your own guide. Paste a house style document and ask for a 400-word draft. Count how many rules it breaks.
- Watch the confirmation behavior. In an agent run on a scratch folder, see whether it actually pauses before deleting or overwriting anything.
- Price December, not October. Budget any recurring workflow at $4 and $20, because that is what the invoice will say after November 30.
The second-opinion routine I already use with Gemini is the natural first place to try this. Cross-checking a Claude draft carries little risk, and it shows quickly whether Argon catches more errors than the model it replaces.
For me, Gemini 4 Argon is the most interesting Gemini release so far because Google led with rollbacks and confirmation prompts, not just scores. I cannot test it today on my current access, so I am treating every number above as Google’s claim until I can. If access opens up before November 30, I will run the five checks on scratch work first. Client work waits until the December price and the real thinking-token bill are both on paper.
Sources
- Google Blog — Gemini 4 Argon: Built to solve your hardest problems
- Google AI for Developers — Gemini API pricing
AI-assisted research and drafting. Reviewed and published by ToolMint.