I spent half a month, Gemini 4 Argon.
Let's start with the results:Gemini 4 Argon is officially open today morning.I spent half a month cruising around for it, stepping on two walls in the middle, writing down below — by the way, where it was won, where it was lost, and showing it to you with an official run schedule.
First wall: The official declaration does not mean it's working.
On September 30, Google launched a blog announcing Gemini 4 Argon, “Our Frontline Model for Real World Programming, Corporate Knowledge Work and Cyber Defense”. It's a scary title, and I was so excited to see how it works.
The last line of the page reads "rolling out soon."Soon was lonely.
We'll find out again. This time it is.First of all, to Fairwind, the "Credible Cyber Security Guard" in the plan.To cooperate with the United States Government in the pre-issuance security assessment. In the words of the product manager, Tulsee Doshi, this was done “to get a model that is well-trained in cybersecurity into the hands of defenders as soon as possible”.
General developers, ordinary users? No way.
Second wall: after 11 days, the network is empty.
I went over it again on October 11th.No API model ID, no model card, no knowledge deadlineI don't know. Those days, the Internet's gone crazy, "Meaning this number," "Crushing that model," but the source, all predictions, not official data.
Here's what I have to say:Google often “shows model names” and “pull models” are different things.On July 21, when the official announced a training, he called Gemini 4, a road map signal, not a serviceable signal.There are many days between the announcement and the service.
It's been 13 days this time.
Three things happened in the morning.
On October 13th (i.e. this morning), Google gave it all:
- Model card release(13 October) — The structure, context and full run schedule are distributed;
- Global openness— Developers, businesses, consumers working together;
- Access path split- Who can use it, from where, in one instance.
I'll put "who can use" in a table, copy it:
| Who are you? | From where? |
|---|---|
| Developer / Enterprise | API (AI Studio + Gemini platform) |
| Google AI Ultra Subscription | Antigravity + Gemini App+ Search AI Mode |
| Google AI Pro Subscription | Gemini App+ Search AI Mode |
| Free. | No official mention. |
Prices were also released:Taste fresh prices $2/ million input token, $10/ million output tokenAnd then it goes up to $4 / $20I don't know. You want to get laid while you're at it.
What is it?
A few words from the official model Kari deserve to be highlighted:
- It's a completely new front-line model, not a fine-tuning or adaptation of the old model.This is important - not "shell upgrading".
- Structure is a large-scale hybrid specialist (MoE) + Transformertext, visual, audio input.
- Context window: Enter + output up to 1M token.This is one of the most substantial upgrades of the generation.
- Training with Google Home TPU, running JAX + ML Pathways.
- Positioning scene:Multi-wheel in-depth investigations, complex engineering tasks, deep visual understanding, cybersecurity defence, multi-mode file analysis, legal/financial/science reasoning.
In one sentence:It's going for the "difficult, deep, long." It's not going for the "go."
It's the top show: the score sheet, the winner and the loser.
This is the most appropriate part of the text. Google DeepMind October 2026The official comparison was GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5.
It won:
| Benchmark | Argon. | GPT-6 | Fable 5.1 | Opus 5.5. |
|---|---|---|---|---|
| Vals Index | 68.9% | 63.1% | 65.8% | 67.0% |
| AutomationBech | 51.3% | 41.4% | 31.4% | 42.5% |
| Vals Finance Agent v2. | 65.4% | 53.5% | 58.9% | 58.6% |
| Harvey, the law. | 19.6% | 5.4% | 6.7% | 3.8% |
| DeepSWE v1.1 (coding) | 77.9% | 74.1% | 67.4% | 74.2% |
| Vibe Code Bench. | 91.9% | 89.6% | 90.3% | 90.3% |
| LABBEC 2 (Scientific) | 88.8% | 85.4% | 68.6% | 73.1% |
| Riemann Bench (Mathematics) | 76.0% | 72.0% | 65.6% | 69.6% |
| GraphWalks 256k~1M | 84.2% | 71.8% | 65.0% | 66.8% |
It didn't win.
| Benchmark | Argon. | GPT-6 | Fable 5.1 | Opus 5.5. |
|---|---|---|---|---|
| FrontierSWE v2 | 55.0% | 65.5% | 56.3% | 62.3% |
| Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% |
| PostTrainBench (ML Project) | 45.3% | 44.3% | 40.2% | 49.3% |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% |
| OSWorld-2.0 (computer operations) | 69.2% | 72.6% | I don't know. | I don't know. |
Cybersecurity, it goes first with GPT-6, Grok 4.7.(CWE-bench v1 is 68.0%).
My sense of reality.
Three directions of real talent:
- The breadth of knowledge work is true.Vals Index, finance, law, and so on -- particularly Harvey Law, Agent, 19.6 per cent against 5.4 per cent, almost crushing. A legal/financial analysis is possible.
- The context is so stable.Graph Walks from 256k to 1M, 84.2% of 71.8% of GPT-6, the difference is clear.You have to feed hundreds of pages of documents to dig deep. This generation is real.
- Multi-mode document analysis is strong.Cartography 71.6 per cent, LVBech 91.7 per cent, which the chart understands as its comfort zone.
The shortboard is also clear:
- It didn't win the hardest real software project.FrontierSWE v2 behind GPT-6 ten percent..."Current" and "hardware engineering" are different things.
- Terminals, scientific reasoning, computer operations are not superior.The person who wants to run a shell task, who wants to control the browser point, is still stronger.
- Don't be fooled by the word "complex first."It's "extremely wide" and not "first of all."
It's on you: part-time advice.
If you're a developer,
- If you want to try, you can start today. AI Studio / API Go on, try fresh prices $2.10;
- But it...Not the answer to a hard-core project like DeepSWE.– You run complex software engineering, start with the gap in the FrontierSWE column;
- 1M context is its most valuable point.Feed it to the long files, to the big warehouse.
If you are a legal / financial / security practitioner
- This generation is for you.A real promotion.(a) Law, Finance, Cybersecurity, first tier;
- Cybersecurity side by side with GPT-6.First distributed to cybersecurity partnersThe move is that Google is seriously betting on the security scene.
If you'd like a taste of it.
- Google AI Ultra, go. Antigravity,Ultra/ Pro subscription in Gemini App and search AI Mode It's in there.
- The free slot is not mentioned.If you want to use it first, there's a pay-off entrance.
Finally.
Google's got it inside.Memory optimization in home-based data centresAnd it says, "Close hundreds of TB RAMs without buying new hardware," and the quantum calculator team is using it. This “self-serving and exterior” route is more indicative than any running score.
But remember today's two walls: the Declarion can be used, and the combination of the first and the first.The lessons of a six-month period are both — the rest, the runsheet says, read it.
It's available.