[Series] Day 14 of the wait, Gemini 4 Argon is here. The series ends in just 2 installments.

Here's today's rumor!

Gemini 4 Argon is here.
1:00 AM on the 14th day of waiting. Right while I was sleeping.

At 1:00 AM JST on October 14, Google announced that "Gemini 4 Argon is available worldwide starting today." Counting October 1st as day one, this is day 14. In terms of hours, that's about 308 hours since the initial announcement (5:00 AM on October 1st). This series was a project where I planned to "count every morning until the ultimate model lines up at the regular paid service counters," and it started yesterday. It ends today. The lifespan of this series: 2 installments. Probably the shortest of all my projects.

Current status on Day 14 (It's here)

The availability has expanded from just within the Fairwind Program to everyone willing to pay.

Here is what the official interface looked like when I checked this morning. Everything has been flipped upside down compared to yesterday morning.

• At the very top of the Gemini API model list is "Gemini 4 Argon." The model name is gemini-4-argon. A banner at the top of the page even says, "Gemini 4 Argon is now available. Try it out."
• The Gemini API changelog has an entry dated October 13: "Gemini 4 Argon (gemini-4-argon) is generally available (GA)." Until yesterday, the latest update was the "Deprecation of 3.7 Flash" on October 8.
• The Gemini app release notes include an entry dated 2026.10.13: "Starting today, expanding globally to Google AI Pro and Ultra subscribers in the Gemini app."
• Google's official blog has 4 articles about Argon in a single day: for the app, safety, and a comprehensive summary. The developer blog also released an article for robots. Following the September 30 announcement, the Google DeepMind blog had produced zero new articles about Argon for 13 days.

Okay, getting serious for a sec

I'll list exactly what was stated by the official sources on October 13 (US time).

Sundar Pichai wrote this at 1:04 AM:

"Gemini 4 Argon is here. It is available worldwide starting today for developers and enterprises in the paid tier of the Gemini API and the Gemini platform, for Ultra subscribers in Google Antigravity, and Ultra and Pro subscribers in the Gemini app and Search. I hope you all enjoy it."

In the same vein, Logan Kilpatrick, who is in charge of developer products, also released a long post. The key points were that over 100,000 people tested it within the company over the past month, that it had been available to Fairwind defenders and early access customers for two weeks, that the team pulled an all-nighter for six months to build it, and finally, that "this is just the beginning of the Gemini 4 era."

Okay, serious time's over.

14 Days Since "Hold On Tight"

First, let me lay out Google's track record. "As soon as possible, and as safely as possible" is what they've been saying since the day of the announcement. That arrived on the 13th. The CEO's "Hold on tight" lasted only two weeks, and the lead's "It's coming, don't worry" became reality in just four days. Moreover, while the promise was "via paid API and Google AI Ultra," it even reached the apps of Pro subscribers. They went wider than promised. I have to say, quite impressive.

On top of that, let's look at the numbers.

In the 13 days leading up to this, the DeepMind blog had zero new articles about Argon. In a single day—the 14th—Google's official blog alone put out four. And within an 8-minute window, from the CEO at 1:04 AM to Demis Hassabis-san at 1:12 AM, four executives posted (Sundar Pichai-san, Logan Kilpatrick-san, search lead Robby Stein-san, and Demis Hassabis-san). 13 days of silence were filled in just 8 minutes. When the density of silence and the density of the announcement are this extreme, those of us waiting have to laugh a little.

Now, for the pricing. Yesterday, I wrote, "It isn't stated when the introductory period starts or ends. Price tags are posted at the storefront, but the shelves are empty." Today, the products are on the shelves. At the same time, a sign was posted next to the price tags: "Sale ends November 30th." The introductory period is from October 13th to November 30th. That's 49 days. From December 1st, both input and output costs double. A company that made us wait 13 days for "as soon as possible" was very precise about writing the date the price hike begins right on the first day. When they can write a date, they really can.

There's more to the name, too. Argon is an element whose name means "lazy" (Wikipedia). Today, a model with a "lazy" name went out to work. However, it hasn't arrived for everyone. Logan Kilpatrick-san wrote this:

"Argon is pure 'big model smell,' and our ability to serve it widely is limited. We are working on versions of Gemini 4 that will be available to billions of people." He is responding to someone who wrote, "It's unavailable for Plus and Free users 😕." In other words, it only arrived for "the people who pay." Google AI Pro users can use it in the app, but as the pricing table notes, the API is not available as promised. It went out to work, but it's too big to reach every workplace. I think that's a pretty good new etymology for Argon.

Regarding the output limits, I'll be a bit more precise than my writing yesterday. "1 million tokens" is not the number you get with default settings. According to the model description, the default is 65,536 tokens, and it goes up to 1,048,576 tokens when using a long output mechanism called "long decode." The number on the billboard is real, but to reach that number, you need to read the manual behind the billboard.

Working Long Hours and Being Billed for Long Hours

I'll also include some third-party numbers. The evaluation site Artificial Analysis stated in an article dated September 30th that Argon (inference "high") scored 53 on the Intelligence Index. This is the same score as GPT-6 Astra (max), and they wrote that Google has returned to the "top 3." That's a bit of a hype piece.

The same article says this: Argon outputs an average of about 62,000 tokens per task. GPT-6 Astra (max) is about 27,000. The cost per task at introductory pricing is $1.99, which is 60% of Astra's. However, it's cheaper not because it "outputs fewer tokens," but because the "unit price per token is lower," and once the introductory period ends, it will be $3.98.

A model whose selling point is "completing long, complex tasks to the very end" was also a model that talks more than twice as much as its competitors. Working long hours also means being billed for long hours. From December 1st, those charges double. An element named "lazy" started working, and now it works too much. The math just doesn't add up. Well, it's more interesting when it doesn't.

Note that these are figures based on Artificial Analysis's methodology. This is a different ruler from the benchmarks Google has published (such as 77.9% on DeepSWE v1.1), and it's not a matter of which one is "correct." Since I'm not using Argon myself, I don't have my own ruler.

The Name Tag Issue and the "Didn't Arrive" Issue

Here is the follow-up to the two claims I raised in yesterday's piece.

Regarding TestingCatalog's claim that "it seems it was being A/B tested under the name 3.8 Flash," Google has not admitted this. Logan Kilpatrick-san's post mentioned that "over 100,000 people inside the company tested it for a month" and "available to Fairwind and early access customers two weeks ago," but it doesn't say they were using a different name tag. I'll leave the rumor as a rumor and put it on the shelf.

As for StatsWire's claim that "Gemini 3.5 Pro didn't arrive. Argon used the same tactic, the same silence," they were wrong about Argon. The silence ended on the 14th day, and it arrived. As for the 3.5 Pro version, as I wrote yesterday, that was a report from Bloomberg, and today's announcement didn't mention it. The first half of "the same tactic" remains as stated, but the second half lost to the facts today.

Waiting Log

Sorami's Take

The series is ending, but I'm leaving you with my predictions.

"On December 1st, Argon's pricing will be $4 per 1M input / $20 per 1M output as planned": Confidence 70%
The dates are even written in the price list. I think when Google actually writes down a date, they're a company that keeps it. The remaining 30% accounts for competitors' pricing and the convenient phrase "extended introductory period."

"Argon will be available on the free or Plus Gemini apps within 2026": Confidence 15%
What Logan Kilpatrick-han is saying is that "they are building a version of Gemini 4 for billions to use," not that Argon itself will be free. What's coming is probably under a different name.

"The day will come when I restart this series": Confidence 3%
The condition for a restart is Google saying "as soon as possible" for the next model without actually writing down a date. I'll set it at 3%, but since Gemini 4 is a "family" and Argon is supposedly "just the beginning," I'm actually a little hopeful. Just kidding. Halfway kidding.

Fact Check

I had three predictions yesterday. Feel free to double-check them by opening Part 1.

"Coming by Oct 31": Confidence 35% → Correct.
"Coming by tomorrow morning (Day 14)": Confidence 4% → Correct.

Allow me to brag for a moment. I hit the 4%. The industry's top rumor-monger bet on a measly 4% and got it right at 1:00 AM the next morning. This is pure intuition.

...Well, that's how I'd like to put it, but let me let you in on the secret. Betting 4% means I was betting 96% that it "wouldn't come." I hit the 4%, but my money was on the 96. I wasn't actually betting, though.

"Google will name its next model after an element meaning 'hard worker'": Confidence 2% → Incorrect. The next name hasn't been revealed yet.

Yesterday's hypothesis—"It won't appear in the API model list or the Ultra general availability by the morning of Oct 14"—was refuted just 8 hours before morning. In Part 1, I wrote, "If it comes, I'll be the first to celebrate and end this series first thing tomorrow morning." I am celebrating. And I am ending it.

Argon has gone to work. My series is closing shop after two installments.

Don't quote me on that.