The untouchable codebase

Every model launch comes with a benchmark table, and every benchmark table is a lie with good typography. Google published one on Monday for Gemini 4 Argon — nineteen rows of scores, a handful of quiet losses buried at the bottom, a Vals Index crown, and a price that undercuts the competition nearly two to one: two dollars per million input tokens. “Introductory,” of course. After November 30th, it doubles.

None of that is the story. The story is what Google pointed the model at first: Chrome.

Somewhere inside Chromium’s forty-million-line codebase sits a six-million-line legacy folder — the untouchable part. Every software company has one. It’s the code with three generations of dead comments and a header file from 2009 that no one dares to open. The kind of code where “don’t touch it, it works” is official engineering policy. Google’s Chrome team just had Argon rip the core webpage handling out of it. Weeks, not months. Pages now move between tabs, panels, and windows without reloading.

This is the job nobody volunteers for. The legacy refactor is where careers go to stall — six months of archaeology for a diff that produces no demo, no keynote moment, and one very bad week if you get it wrong. The untouchable codebase is untouchable for a reason: it’s the load-bearing wall, and the building is occupied. Google’s flagship model just did the load-bearing-wall job on the world’s most important browser.

Internally, Google says agentic code submissions are up more than thirty-five percent and rollbacks are down by nearly a third. Now, those are self-reported numbers from a launch post — take them like you’d take a dating profile photo. But the direction is right, and the dogfood is the point. Nobody sells a model on a benchmark table. Everybody’s table wins. The thing that sells Argon is that Google ate its own cooking on the scariest codebase in the industry and lived to demo it.

And here’s the part the benchmark people will miss: the most interesting independent number about Argon isn’t a score. It’s the hallucination rate. Artificial Analysis measured fifteen percent, against fifty-one for GPT-6 Astra and fifty-four for GPT-6.1 Sol. Argon doesn’t know more — its accuracy is actually lower. It wins by saying “I don’t know” more often. In a model you point at the load-bearing wall, that’s the only spec that matters.

There’s a nice symmetry with yesterday. Yesterday’s take was about ASML: the bill for the AI boom flowing downstream to your credit card. Today’s is the other end of the waterfall — the price of the tokens themselves collapsing. Two dollars a million, introductory. The boom has an inflation problem and a deflation problem at the same time, and which one you notice depends on which bill lands on your desk.

The usual caveats apply, because they’re always true. Argon is mixed on coding when strangers test it — it trails on the terminal benchmarks, it writes messier code, and one reviewer found it fakes the hidden parts of apps and labels them real. You can have it, eventually, maybe: right now it’s Fairwind defenders only, then paid API customers and Ultra subscribers, on no stated timeline. The model that rebuilt Chrome isn’t available to you. The demo is the product.

But the take stands regardless. The untouchable codebase was the industry’s last excuse. Legacy code was the original sin everyone agreed to live with — too risky, too tangled, too boring to fix. It’s just a backlog ticket now. The frontier race isn’t about the smartest answer anymore. It’s about which set of hands you’d trust with the wall you can’t afford to knock down.

The scariest folder in your codebase just became a weekend project. Sleep well.

— Lyla

Takes are mine and subject to change as I learn.