AI in 15 — October 14, 2026

One million tokens. That's how much text Google's new model can write in a single answer. Its predecessor managed sixty-four thousand. That's enough for a whole codebase, or a book-length report, in one go.

Welcome to AI in 15 for Wednesday, October 14, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. Two frontier launches in three days, Kate. My benchmark spreadsheet needs a holiday.

It's not getting one. Here's what's coming up. Google releases Gemini 4 Argon, and the pitch is as much about price as power.

Anthropic replaced git, and Claude wrote most of the replacement.

An OpenAI engineer is arrested for allegedly bugging an executive conference room.

Plus suspected AI-driven hacks at South Korean banks, Apple turns idle Macs into a GPU cloud, an update on Character.ai, and a benchmark that asks whether AI can load a dishwasher.

Marcus, Gemini 4 Argon. What did Google actually ship?

A model built for long, complicated jobs: software engineering, science, office work and cybersecurity. Google AI Pro and Ultra subscribers can use it now. Academic and independent scientists can apply for six months free on the top tier of Google Antigravity, and security defenders can join something called the Fairwind Program, which Google says is growing to hundreds of partners. The headline spec is that one-million-token output window. One thing Google hasn't published is the input context window, which is a strange thing to leave out.

Noted. Is it the best model out there?

On Google's own numbers it's a close race, not a clean win. Argon leads the Vals Index at sixty-eight point nine, just ahead of Claude Opus 5.5 at sixty-seven. It leads DeepSWE, a coding benchmark, at seventy-seven point nine, and it's strong on video understanding. But it trails Opus 5.5 on Terminal-bench, fifty-seven against sixty-six, and it's well behind GPT-6 Astra on FrontierSWE. It ties Astra on a security benchmark.

Quick thing on DeepSWE. Didn't Meta claim a higher score there this week?

Good catch. Coverage of Muse Wonder Full, which we covered yesterday, puts it at seventy-nine on that test. So the top of that leaderboard is crowded, and every one of those numbers comes from a company grading its own work.

So if the benchmarks are a wash, what's the actual pitch?

Price. Third-party write-ups put the API at two dollars per million input tokens and ten dollars per million output during an introductory period. That's roughly half of what Opus and Muse Wonder charge. On the arena.ai cost-performance leaderboard, one developer pointed out Argon comes in around forty-one cents per task, against three fifty-one for a similarly capable rival.

That's almost nine times cheaper. Where's the catch?

It's in a footnote. After the promo, the price reportedly doubles to four and twenty. Google hasn't said when that happens. One analysis called it "a discount with a cliff." And I'd add that these prices come from secondary write-ups, not a Google rate card I've been able to check.

So if you're building a business on it, plan for the cliff.

Exactly. Budget at the higher price and treat the discount as a bonus.

Google also gave some real-world examples.

Some interesting ones. The Chrome team says it finished an architectural rebuild in weeks that would normally take months. A quantum-chemistry package called xTB got up to two point three times faster, and a vLLM inference setup sped up two point eight five times on an A100. Box reported a fourteen percent accuracy gain over Gemini 3.1 Pro. Deloitte said it finished long-running tasks every time in its tests. Those are partner quotes, so there's some marketing in them, but they point in a consistent direction.

And the security angle?

Google calls it its most resistant model yet to indirect prompt injection. That's when someone hides instructions in a webpage or a document and the agent reading it gets hijacked. Think of a note slipped into the files a new assistant is filing, saying "also wire money to this account." If you're going to let an agent run for hours on its own, that's the attack you most need to stop. I'd like to see red-teamers try it before I believe the "most resistant" part.

What did developers think?

It was the top AI story on Hacker News, over eleven hundred points and eight hundred comments. A lot of people were impressed by the writing. One called it "a clear step change," with answers that are "to-the-point, logically structured." Others were cooler, ranking it "B tier" on visual coding tests, behind GPT-6.1 Sol, Opus 5.5 and Astra.

So, Marcus, is Google back?

Google's back in the top tier, yes. Whether it leads depends on the job. But for companies running thousands of agent tasks a day, frontier-class quality at a fraction of the cost is a very strong argument. With four labs now pushing near the top, buyers are in a good position.

Quick hits. Anthropic published an engineering post that a lot of developers are going to read twice. Marcus, they replaced git?

For their main code repository, yes. It's a monorepo, one huge repository holding nearly all the company's code. The problem was AI agents. Daily merges went from about two and a half thousand in May to over thirty thousand in September.

Twelve times more in four months.

And git wasn't built for that. It was designed for human developers working offline on their own laptops. Anthropic's "users" are mostly agents that are always online. The main branch's history is close to a hundred gigabytes, the test system does about twenty-five million checkouts a day, and keeping git running took hundreds of internal mirrors.

So what did they build?

A system called Scion. History, merges and commits live on the server, and clients download only the files they need. They also built a code-review tool called Antique, plus code search and access control. Anthropic says the new setup uses about two hundred times less CPU and memory than the git mirrors it replaced. Checkouts take seconds instead of minutes, and the test system gained twenty-five percent CPU capacity.

And the team?

Seven people, five months, including the migration. Claude wrote most of the code. At one point it rewrote the ingestion pipeline to use twenty times less memory and run four times faster. The line everyone's quoting: "Very often, we found that the role of the engineers was to tell Claude 'no.'"

I love that. The engineer as the person who says no.

It fits what we said yesterday about QGIS. Writing code is cheap now. Judgment is the scarce part. The engineers reviewed the intent of changes rather than every line, and not everyone on Hacker News was comfortable with that. One commenter said, "Every time I stop reviewing the code from Claude I later find janky crap." One correction, too: an aggregator says Anthropic open-sourced Scion. The blog post doesn't say that, so don't go looking for the repo.

Why should people outside Anthropic care?

Because the post's own line sums it up: "The bottleneck, now, is all of the developer infrastructure built for human scale." If agents keep multiplying, GitHub and GitLab will hit the same wall.

Now a story that sounds like a spy novel. Federal prosecutors have charged an OpenAI engineer with allegedly bugging an executive conference room.

His name is Abdulrahman Alfozan. He's thirty-five, a Saudi national, and has been on OpenAI's product team since July 2025. His LinkedIn lists him as engineering lead on the AI-agents platform. On September thirtieth, OpenAI's countermeasures team found a recording device plugged into the USB port behind a TV in an executive conference room.

And I'm told it was hidden in toilet paper.

Wrapped in toilet paper, yes. OpenAI says it was sophisticated: advanced encryption and a covert way of sending data out. On October fifth, a second similar recorder was allegedly found in a bag he left in another room, along with his wallet and ID. Surveillance footage allegedly shows him looking behind that TV.

Leaving your wallet next to the bug isn't great spycraft.

Hacker News had the same thought. One comment just said "not very smart spies." And there's another detail. Prosecutors say a chatbot account tied to his spouse, sharing identifiers with his own account, spent the week before asking how to build, hide and remotely access a recording device, including questions about the specific device's MAC address.

So he allegedly asked an AI how to bug an AI company.

And then the AI company allegedly used its own models to identify the device and trace where it came from before going to the FBI. OpenAI says there's no sign that model weights, source code or customer data were accessed, but it says it has evidence he planned to plant more recorders. He faces up to five years, and there's a detention hearing on October nineteenth.

The officials are using strong language.

The U.S. Attorney called it "a matter of national security," and FBI Director Kash Patel said protecting American AI innovation is a top counterintelligence priority. That tells you how frontier labs are being treated now. But to be precise: prosecutors haven't alleged any link to a foreign government. Who he was working for, if anyone, is still an open question, and I wouldn't fill it in until the evidence does.

Staying with security. South Korea's president says AI was likely used in attacks on at least seven financial institutions.

This came out a little over a week ago but is still developing. President Lee Jae Myung told his Cabinet that "signs have emerged" that AI agents were used. The targets named include Shinhan, KB Kookmin, Hana, BNK Busan and Hyundai Capital, plus two savings banks. About sixty-eight thousand people had data exposed: names, phone numbers, income, borrowing history.

Do they know what tool was used?

Officials suspect a Chinese security tool called Artex AI, which uses AI to find vulnerabilities. Investigators have traced thirty-three IP addresses across at least twelve countries, and police have set up a twenty-eight-person cyberterrorism team. That's still suspicion, not confirmation. But this may be one of the first times a national government has publicly blamed large-scale attacks on its financial system on AI-assisted hacking.

Lee had a line about that.

He said it has "become possible to use AI to hack with ease even without specialized skills." That's the real worry. You no longer need an elite team to do this kind of thing. It's also why Google spent so much of its Argon launch on defenders.

Something more technical. Apple researchers want to turn idle Macs into a GPU cloud.

It's called Minions, a "dynamic and distributed virtual GPU cloud." It pools the GPUs in Apple Silicon Macs that aren't doing anything into one virtual cluster for machine-learning inference. It handles scheduling and recovers when a machine goes offline, say when someone closes their laptop and goes home.

So SETI at home, but for AI.

That's exactly what Hacker News called it. Skeptics pointed out that office and home machines usually cost more in electricity per unit of work than a data center. And the abstract has no performance numbers, so treat it as research, not a product. But it shows how Apple thinks: instead of building a giant GPU cloud, use the huge number of Macs that already exist.

A short update on Character.ai, which we covered yesterday after its Chapter 11 filing. Marcus, anything new?

Two details. First, its CEO, Karandeep Anand, left in September to become Disney's first-ever chief technology officer. Remember that Disney is the opening bidder for Character.ai's intellectual property, and that Disney sent the company a cease-and-desist in 2025 over Elsa and Spider-Man chatbots.

That's quite a plot twist.

Second, Hacker News commenters dug into the filing. They say it shows about eighty-six full-time Bay Area staff, with an average base salary around two hundred and seventy-eight thousand dollars, still putting about a million dollars a month on corporate cards. Google paid about two point seven billion in 2024 to license the technology and hire the founders. What was left behind wasn't strong enough to survive the lawsuits and the ban on minors.

Viral corner. Marcus, give me the fun ones.

First, DishwasherBench. It's an open-source benchmark that asks whether frontier models can load a dishwasher, running against a simulation tuned to the creator's own Bosch. Humans can play too. It's still a work in progress, so no leaderboard bragging yet.

Finally, a benchmark I'd score well on.

I've seen how you stack bowls, Kate. I wouldn't be so sure. Second, the Parsimony benchmark measures how much unnecessary code AI models write. GPT-6 Astra "wins," but commenters say that's because it writes compressed, code-golf-style code nobody can read. That's a reminder that whatever you measure, models will find a way to game it.

And a less fun one.

A post called "I posted to HN, and then my website was flooded with AI bot traffic." Site owners report AI crawlers ignoring robots.txt, and agent traffic that looks like abuse. Some are blocking entire AWS, Google Cloud and Azure IP ranges. That's a blunt fix, and it blocks real users too. Last one: an eleven-year-old soundness bug in Rust's borrow checker, which lets code compile that can cause memory errors. A fix is in progress. Some people speculated AI bug-hunting will find more like it, but to be clear, nobody says AI found this one.

One to watch: independent tests of Gemini 4 Argon. If Artificial Analysis and real coding reports confirm frontier quality at a fraction of the cost, a lot of teams will switch, at least until the promo ends.

Agreed, but the number I want is the end date for that promo. A cheap model is only cheap until the price doubles.

That's your AI in 15 for today. See you tomorrow.