From the separate room to the classroom at last. Gemi-chan gains the power of "Argon".
Meet Gemi-chan (Google)'s top model, Geminiย 4 Argon. She was announced on September 30, but went first only to cyber-defense organizations โ a "separate room." Then on October 13, 2026 (U.S. time), she finally joined the class through paid plans and the API. What's new, who can use her, and how she differs from Kuroko and the others โ gently explained.
So what's actually new in Geminiย 4 Argon?
The biggest change: the "4" generation was built from scratch. The name follows "generation (4) + name (Argon)". Argon is spelled like the chemical element, though Google hasn't explained the choice.
A brand-new model
The model card describes it as a new frontier model, not a modification or fine-tune of an existing one. Under the hood it's a large sparse MoE (it calls up only the "experts" it needs), trained on Google's TPUs.
Bigger than the old Pro
A Google spokesperson said it is bigger than the previous Pro models. The planned Geminiย 3.5 Pro release was called off (Reuters, Oct 1).
Writes up to a million tokens
Input and output share 1,048,576 tokens (about 1 million). The September 30 announcement said the output limit rose from 64K to 1M. In the API, a standard request defaults to 65,536 output tokens; with long decode it can go up to about 1 million.
What she reads and writes
Input: text, images, video, audio, and PDF. Output: text only. According to the model card, her knowledge runs through March 2026.
Where can you use her?
As of the October 13, 2026 announcements (Google Cloud: October 14).
Google says it is rolling out globally, so in Japan, too, it may take a while before she appears in the model picker. There is no Japan-specific announcement.
| Who | Where she's available |
|---|---|
| Free plan | Not included |
| Google AI Pro | Gemini app, AI Mode in Search |
| Google AI Ultra | All of the above + Google AI Studio and Google Antigravity (a coding tool) |
| Developers | Paid tier of the Gemini API (model ID gemini-4-argon). No free tier |
| Businesses | Gemini for business, Google Cloud |
| Cyber-defense organizations | A version with cyber-related restrictions lifted, rolling out via the Fairwind Program. On Oct 13, expansion to hundreds of additional partners began |
| Researchers | 10,000 vetted scientists get Antigravity's top tier free for six months |
For API pricing (per million tokens), the key is the introductory price through November 30.
| Model | Input | Output |
|---|---|---|
| Geminiย 4 Argon (intro price, through 11/30) | $2 | $10 |
| Geminiย 4 Argon (from 12/1) | $4 | $20 |
| Claude Opusย 5.5 (Kuroko) | $4 | $20 |
| GPT-6ย Astra (Chappy) | $10 | $50 |
Batch / Flex cost half. Once the regular price kicks in on 12/1, she costs the same as Kuroko's Opusย 5.5.
- In the Gemini app, answers come with less preamble and repetition.
- You can now switch between chat and tasks.
- The macOS app gets Antigravity's coding view.
Gemi-chan's strongest subjects
All figures are Google's own, as reported in its model card. The numbers in parentheses are, in order, GPT-6ย Astra (Chappy) / Claude Fableย 5.1 / Claude Opusย 5.5 (both from Kuroko's family).
Work automation
AutomationBench 51.3 (41.4 / 31.4 / 42.5). A test of how well a model carries out business workflows on its own.
Law
Harvey Legal Agent 19.6 (5.4 / 6.7 / 3.8). Among those compared, she's far ahead of the rest.
Long coding tasks
DeepSWE v1.1 77.9 (74.1 / 67.4 / 74.2). Internally, Google says splitting up an old Chrome folder of about 6 million lines went from months to weeks.
Overall
Vals Index 68.9 (63.1 / 65.8 / 67.0). An overall score that measures many kinds of work together.
Understanding long videos
LVBench 91.7 (87.5 / 79.7 / 83.7). A very Gemi-chan subject โ she can read video.
Hard to trick with slipped-in instructions
Against instructions slipped in from outside (indirect prompt injection), the attack success rate on Gray Swan's test is 0.7%. According to VentureBeat, Opusย 5.5 scored 1.0% and Astraย 8.5%.
She has weak subjects, too. Computer use on OSWorld-2.0: 69.2 (Astraย 72.6). Terminal work on Terminal-bench 4.0: 57.4 (Opusย 5.5ย 66.4). FrontierSWE v2: 55.0 (Astraย 65.5). By VentureBeat's count, of the 18 published benchmarks she took 12 outright firsts and 1 tie for first. Reuters, however, noted that the figures are Google's own and that she trails on two of the four coding benchmarks.
Why two weeks in a "separate room"?
When she was announced on September 30, Argon was available only to cyber-defense organizations in the Fairwind Program and inside Google. She also went through the U.S. government's voluntary pre-deployment testing.
Defenders go first
The Fairwind Program started on September 2 for the "defenders" of governments, critical infrastructure (healthcare, telecom, energy, finance), and major tech platforms. It has more than 650 partners โ but that's the program as a whole, not the number of organizations that received Argon.
The reason lies in the model card's safety evaluation. Under Google's Frontier Safety Framework, biology and part of cyber could not be ruled out as having reached a dangerous threshold. Google says its safeguards bring the risk back below that threshold. Chemical, radiological and nuclear risks, harmful manipulation, and ML research and development did not reach the threshold.
The safeguards come in three parts: refusing harmful requests while still helping legitimate research / monitoring internal behavior / internal and external red teams. External red teams got past the safeguards on less than 1% of the biology subtasks. The general-release Argon blocks offensive cyber work such as penetration testing and writing exploits, and her reasoning and actions are monitored in real time. Google also says attempts to "cheat" on coding evaluations fell to 2%.
From announcement to first day of school
From the defenders' head start to the end of the introductory price, in order.
Sep 2 โ Fairwind begins
The Fairwind Program for defenders starts, initially for Geminiย 3.8 Flash Cyber.
Sep 30 โ Geminiย 4 Argon announced
But only as a limited release for cyber-defense organizations.
Oct 1 โ 3.5 Pro called off
Reuters reports that the Geminiย 3.5 Pro release has been canceled.
Oct 13 โ Into the classroom
Opens to Pro and Ultra subscribers and the paid API. Fairwind also begins expanding to hundreds of partners.
Oct 14 โ Google Cloud
Now available on Google Cloud for businesses, too.
Nov 30 โ Intro price ends
From Dec 1, the API costs $4 in / $20 out.
Report cards: Gemi-chan vs. Kuroko
Two students who will cost the same ($4 / $20) from December 1, side by side on the independent Artificial Analysis Intelligence Index (the version as of October 2026) โ
๐ Still ahead overall
- Overall Intelligence Index 58
- Her younger sister Sonnetย 5.5 follows at 56
- Price: $4 in / $20 out
๐ Tied with Astra-chan
- Overall 53 (High setting) โ tied with GPT-6ย Astra and Fableย 5.1
- Cost per task at the intro price: $1.99 (Astra: $3.26). At the regular price it becomes $3.98, about 20% more than Astra
- The reason: long answers. About 62K output tokens per task โ more than twice Astra's (about 27K)
- A low hallucination rate of 15%, but 50% accuracy on knowledge questions, below Astra's 63% (The Decoder)
The Decoder's verdict: she "closes the gap, but doesn't take a clear lead." On the Vals Index, though, her 68.9 is the highest of the four compared in Chapter 3. Note that the scores in our Astra feature (such as Fableย 5.1's 66) come from a different version of the index, so the numbers can't be compared directly. "Kuroko is still ahead overall, and Gemi-chan has drawn level with Astra-chan" is the honest state of things as of October 2026.
Gemi-chan, too, came to school โ with a separate room attached.
Not "stop her," not "release everything," but hand her to defenders first, firm up the safeguards, then let her into the classroom โ following Astra, Gemi-chan was released the same way. On top of that, the cheap introductory price through November 30 is a play to get people using her fast. Google says "Argon is just the beginning" and promises the rest of the Geminiย 4 family, but no timing or names have been announced yet.