News of the day. On October 13, 2026, Google opened Gemini 4 Argon to all paid users. The model is live in the Gemini app, in AI Mode in Search, and through the API. On real work tasks it beats Claude Opus. On price, it costs half as much. But do not trust benchmarks. Test it on your own task. It takes 10 minutes.
The short version
- Live since October 13. Argon rolled out to all paid users: Google AI Pro and Ultra, the Gemini app, AI Mode in Search, and the Gemini API for developers and business.
- API price: $2 / $10 per 1M tokens — $2 input, $10 output. Claude Opus 5.5 is $4 / $20. Exactly twice as much.
- Ahead on work tasks. AutomationBench: 51.3% for Argon vs. 42.5% for Opus. Full numbers below, each with its source.
What happened
The announcement came on September 30, with limited access at first. On October 13, Google opened the model fully — to all paid plans and to the API.
One clarification matters: there is no free access. "Everyone" means everyone who pays — Pro, Ultra, or paid API. Antigravity is Ultra only.
The release itself is not the main point. Where the model already works is. Argon answers your clients in AI Mode in Google Search. A client asks about your service, and this model writes the answer. Whether it names you or a competitor is now a question of your visibility. Not of luck.
The numbers
- AutomationBench (Zapier): Argon — 51.3%, Claude Opus 5.5 — 42.5%, GPT-6 Astra — 41.4%. This test measures automation work, not trivia.
- DeepSWE v1.1 — 77.9%, according to Google.
- Vals Index — 68.9%, ranked #1, according to Google.
- Context: a shared limit of 1,048,576 tokens, with output up to roughly 1M tokens on long requests, according to Google.
The price comes with one condition. $2 / $10 is the introductory rate. It runs until November 30, 2026. After that, the standard rate is $4 / $20. Cached input costs 95% less: $0.10 per 1M tokens.
So the picture is simple. During the introductory window, Argon costs half of Opus — and scores higher on work benchmarks. Windows like this do not stay open long. Test it now, not in December.
The main part: test the model on your own task
A benchmark does not pay you. A closed task does. Here is a three-step test. One real task of yours is enough.
Step 1. Open the model
Open the Gemini app. Select Argon in the model picker.
No subscription? Open Google Search in AI Mode. The model is already there.
Step 2. Give it one real task
Not a sample question from the internet. Your work. For example, a commercial proposal based on your price list.
Copy this prompt exactly:
Put together a commercial proposal based on this price list. Format: table. Item. Price. Timeline.
Paste in your price list. Start a timer.
The readiness test is simple: the result can go to the client with no edits. If it can — the task is closed. If it cannot — step 3 shows you where it broke.
Step 3. Check the result at three points
Three points. That is all you need at the start:
- Numbers. Every price and every total, checked against your price list.
- Names. Item names, service names, people's names — nothing rewritten, nothing invented.
- Timelines. The timelines from your price list, not the model's rough guesses.
A mistake in the numbers means one thing: the model drafts only. Nothing like that goes to a client.
All clean — and at least a third faster than your manual work? Then hand the task to the model permanently. As a role, not as a one-off experiment.
Bonus. Check who AI Mode names
Ask AI Mode about your service in your city. Word it the way a client words it.
It names a competitor, not you? That is no longer a question for the model. It is a question about your visibility in AI search. The model simply showed you the truth: it recommends whoever it knows in your niche.
The takeaway
Strong AI got cheaper and available to everyone paid. That is the fact of this release.
Access is now the same for all. The winner is not the one who has the model. The winner is the one who put it on their own tasks first — and checked the output by numbers, names, and timelines.
What to do next
Two actions for today. First: run the three-step test on one of your tasks. Ten minutes, and you have your own answer instead of someone else's benchmark. Second: ask AI Mode about your service in your city and see who it names.
I have been building websites and advertising since 2009, and for the last 3 years — artificial intelligence. I build full multi-agent systems that replace entire departments for businesses and experts. If AI Mode names someone else for your service, that is a visibility task — and it can be solved systematically.
If you want to solve tasks like this systematically, start with the service page:
AI agent systems for business →
*By Ievgen Kovalevskyi, AI automation architect.*
Sources