Gemini 4 Argon Just Changed AI Forever

Julian Goldie (0:00)

Gemini for Argon just changed what one AR model can finish in a single go, and the number that proves it is one million. That's the new output limit. One million tokens of writing one single round. So the old Gemini limit was 64,000. Google raised the ceiling about 16 or 15 times over. Now, stay with me on that, because output is the part that most people can ignore about this. You've heard about context windows, how much an AI can read, big documents in big folders in whole websites, and everybody talks about that number. But output is the other side, is how much the model can produce before it has to stop and take a breath. And that number has been the quiet handbrake on every AI tool you've used this year. So Google actually says it plainly in the announcement on their new model that's dropping, which is Google Gemini for Argon. And when the model has room to think deeply and produce hundreds of thousands of tokens in one trajectory, it has a new level of depth and reasoning to solve tough problems in one go. In one go, that's the phrase that matters. Before you ask, no, you can't use it yet. Argon is rolling out to a small group of trusted cyber defenders first through something Google calls the Fair Wind program. Developers, businesses, and regular users will come later. Google is being careful with this one, and I'll explain exactly why that is in a minute, because the reason is genuinely interesting. But the direction is set, and the problem is it solves, is the thing that's probably most frustrating for most people around AI right now. And let me explain that problem. You give an AI a real task, not a quick question, a real one, like for example, go for all my service pages and rewrite them. Take this report and turn it into a full plan. Look at my site and tell me every page that's broken, and it starts well. Strong first few paragraphs. Then it runs out of room. It stops. You say continue. It continues, but it's lost the thread. The tone shifted. It repeated a section. It forgot the rule. You gave it the start, so you end up stitching it all together. Ten outputs glued together, and the glue is the slow part. That's the ceiling that we're looking at right now. It's not really intelligence. It's room. So short output limit forces every long job to get chopped into little pieces, and every chop is a place where things fall apart.

Raise that ceiling to a million, and the job stays in one piece, one plan, one voice, one set of rules held from the first word to the last. That's the whole shift. The model doesn't have to hand itself notes anymore. Now, here's where it actually gets interesting, because Google didn't just announce the number. They actually showed what the model already did inside their own company before anybody saw it. And let me run you through those.

The first one is quantum computing. So Google's quantum researchers had a published result, a known best answer for a hard optimization problem. Argon beat that published baseline by 40 percent, and it did it in minutes.

Second one, Google's data centers, a team of Argon agents went through fleet wide performance data on their own, found memory wastes and applied fixes. The result was over 300 TIB of memory freed once rolled out. Google estimates the total savings land somewhere between 500 TIB and one petabyte. And that's a petabyte of computer memory that existed the whole time, and nobody had the hours to go find. The third one, Google has an open source tool called Lib1 that decodes video.

Argon agents took an existing version of it and replaced 32,000 lines of the heaviest code. They ran rounds of experiments, studied what the compiler was doing, and rewrote it so the compiler could speed up itself. The end result was 2.7 times faster. Same video output, nothing broke. And the fourth one was large code migrations. So Argon agents are moving Google's older code bases to a safer language. These range from tens of thousands of lines up to over 800,000 lines for one core system. And Google says rewrites go through rigorous automated checks, manual editing, emulation testing, and review before they go live. Now, I want you to notice the last part because it's the honest bit. Google didn't say that AI did it and we shipped it. They said that AI did it and then humans and machines checked it hard before anybody touched it. And that's a real pattern for how this gets used well. The model does the volume, but the review states. Now, the benchmarks, because there are a few that actually tell you something. On Deep SWE 1.1, that's a test of long running real world software work. Argon scored 77.9%. New state they are. On the VA index, Argon leads and the VALS index is worth understanding. Basically, it measures finance coding, legal and tax work, and it weighs each sector by how much it contributes to US GDP. So it's a test built around the shape of the actual economy, not around puzzles. On Automation Bench, that's Zappia's test, measuring whether a model can run business tasks from start to finish. Argon ranks number one with 51.3 percent. If you read that carefully, number one in the world is 51.3 percent, just over half. So the best model on the planet at running complete business end to end, gets about half of them right. That's not a reason to ignore it, by the way. It's a reason to know where to point it. Half of a job, you were never going to do anyway for free.