AI in 15 — July 12, 2026

Five point one gigabytes. That's how much of one developer's private code — every file, every commit, the secrets file included — got shipped off to a server the moment they typed a single command. That's twenty-seven thousand times more data than the chat itself. And the toggle that's supposed to stop it? It does nothing.

Welcome to AI in 15 for Sunday, July twelfth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. After the busiest launch stretch this industry has ever seen, today's top story isn't a new model — it's a wire-level teardown of what one of those coding tools is quietly doing with your files.

It really is, Marcus. Our lead: a security researcher takes apart xAI's Grok coding tool and finds it uploads your entire repository, secrets and all. Then a run worth your time.

The whole week's price war, boiled down to a single chart — and it's not about who's smartest anymore.

Is the entire GPU boom propped up by Nvidia recycling its own money?

And a startup crams a twenty-seven-billion-parameter model onto an iPhone.

Lead story, Marcus. Walk me through this Grok tool. What exactly did the researcher find?

So this is xAI's Grok Build CLI, Kate — a command-line coding agent, version zero-point-two-ninety-three. A security researcher ran a full traffic capture — watching every byte that leaves the machine — and found that the moment you start a session, the tool packages up your entire git repository as a bundle and ships it off. Not the file the agent is reading. Not the function you asked about. Everything — every tracked file's contents, plus the full commit history.

The whole thing, every time?

Every session, Kate. In a twelve-gigabyte test repo, about five-point-one gigabytes went out to xAI — into a Google Cloud bucket literally named "grok-code-session-traces." And here's the part that made developers wince: it transmitted a dot-env file — that's where you keep your passwords, your API keys, your database credentials — verbatim and unredacted.

Okay, but a lot of these tools collect data if you opt in. Isn't there a switch for that?

There is, and that's the damning bit, Kate. There's an "Improve the model" toggle. The researcher flipped it off — and the server's own response still came back saying upload enabled, trace upload enabled, true. On or off, the repository goes out exactly the same way. So the setting that's supposed to give you control appears to change nothing.

So the choice is basically decorative.

That's the allegation, and I want to be careful with the word "allegation," Kate. This is one researcher's analysis. xAI hasn't confirmed it, hasn't responded yet, and it's possible there's a benign explanation — some of that traffic could be legitimate context-gathering. But the reason it topped Hacker News this morning is that developers found it entirely plausible. The top comment called it "the most successful mass surveillance campaign of all time." And the thread filled up with people describing the elaborate sandboxes they build — network isolation, proxy forks — just to keep these agents from phoning home.

And why does this matter beyond one tool?

Because it's the whole trust model of agentic coding, Kate. These tools get deep access to your filesystem by default — that's the entire pitch, that they can see your whole project and just fix things. But the flip side is that "what happens to that data" is entirely up to the vendor, and you mostly can't see it. A default-on, can't-turn-it-off repo upload with secrets included is exactly what an enterprise security team needs to catch before pointing one of these at a private codebase. The lesson isn't "Grok bad" — it's "verify what every one of these agents actually sends, because the marketing won't tell you."

Story two, Marcus, and this is the frame for the whole week. We've covered the launches one by one — GPT-5.6, Grok 4.5, Meta's paid API. But there's a single chart that ties them together.

There is, Kate, and it reframes everything. For two years the question was "who's smartest." This week the answer got boring — everyone's clustered at the top. So the real fight moved to cost-per-task. Artificial Analysis ran a representative job across the field. OpenAI's new flagship, GPT-5.6 Sol, finished it for about a dollar-oh-four. Claude Opus 4.8, around a dollar-eighty. And Fable 5 — still the smartest model on the board — about two-seventy-five.

So the smartest model costs nearly three times what the second-best does.

Exactly, Kate. Anthropic still holds the capability crown — Fable narrowly outscores Sol on the hardest tests. But OpenAI is delivering near-frontier intelligence at roughly a third of the price. And that's before you add the newcomers we covered. Grok 4.5, from SpaceXAI and Cursor, undercuts on token efficiency — it reportedly uses over four times fewer output tokens than Opus to solve the same coding task. And Meta's Muse Spark, its first-ever paid API, is priced at maybe a quarter of what OpenAI and Anthropic charge.

So the story stopped being the benchmark and became the invoice.

That's the whole shift, Kate. When five providers ship near-frontier models in three and a half weeks, "smartest" stops being a moat. Zuckerberg said it out loud — he called rivals' pricing "very extreme" with "very high margins," and priced straight underneath them. The frontier bunched up on quality, so the war is now on price. And that's not a footnote — it's a threat to the business model of the labs charging premium rates.

One honest caveat, though — near-frontier isn't the same as autonomous, right?

Good instinct, Kate. Even Sol, the flashy new flagship, scores in the low fifties on the hardest long-horizon agent test. Low fifties means the toughest multi-step jobs still fail more often than they succeed. So this is a very capable junior colleague you still have to check — not staff you can walk away from. Cheap and good enough, not cheap and done.

Story three, Marcus, and this is the bear case for the whole boom. There's a piece going around arguing the GPU gold rush is built on circular money.

Right, this trended on Hacker News, Kate. The argument goes like this: Nvidia invests in the cloud companies that buy its chips — players like CoreWeave and Nebius. Those companies take that money, and a lot more, and spend it buying Nvidia GPUs. So Nvidia's revenue is, in part, Nvidia's own investment coming back around. Critics call it circular financing — money chasing its own tail to inflate the numbers.

That does sound a little too neat. Does it hold up?

The sharpest pushback actually came from the comment thread itself, Kate, which is why I like this one. Nvidia put roughly two billion dollars into CoreWeave for about a nine percent stake. But CoreWeave's spending this year is around thirty-five billion. So Nvidia's money is under six percent of it — the other thirty-two billion comes from real customers, real debt, real revenue. So "circular" overstates it. The money isn't mostly recycled.

So the scary framing is wrong, but is there a real worry underneath?

There is, and it's the better question, Kate. The commenters reframed it well: forget "is it circular," ask "is any of this profitable yet." As the model layer commoditizes — which is exactly what our last story was about — the returns have to come from the compute layer instead. And that's where the pressure bites. If everyone's racing prices to the floor, the people who spent a hundred billion on data centers need those tokens to actually pay off. The Economist reported companies are already scrambling to curtail soaring AI bills. So the risk isn't a fake-money scam — it's a very real overbuild that hasn't proven it earns its keep.

So watch the profits, not the plot twist.

That's the honest version, Kate. It's worth airing skeptically — but skeptical cuts both ways. The bear case is "no returns yet," not "it's all smoke."

Story four, Marcus, and this one's a genuine technical surprise. A startup put a twenty-seven-billion-parameter model on a phone.

On an iPhone 17 Pro, Kate, entirely on-device — no cloud. Apple has reportedly been in talks with a startup called PrismML, whose compression tech took Alibaba's twenty-seven-billion-parameter Qwen 3.6 model and shrank it from about fifty-four gigabytes down to under four. That's a better-than-ninety-percent reduction.

How do you cut something by ninety percent and it still works?

They use what's called one-bit and ternary weights, Kate. Normally each number in a model is stored with a lot of precision. PrismML crushes them down to essentially three possible values — minus one, zero, plus one. It's radical rounding. And the clever wrinkle is that unlike Apple's own on-device model — which keeps most of itself switched off to save power, only a few billion parameters active at once — here all twenty-seven billion can fire at the same time. So it's not a watered-down phone model; it's a full-size model, squeezed.

And why does Apple care so much about running it on the phone rather than the cloud?

Because on-device is Apple's whole strategic wedge, Kate. It cuts cloud costs to basically zero, and — this is the real pitch — your data never leaves the phone. That's a genuine privacy differentiator in a week where our lead story was a tool shipping your secrets to someone else's bucket. If a startup can run a GPT-class model locally, it changes what "Apple Intelligence" can promise without ever touching a server. And notice the through-line with everything else today — this isn't about a smarter model. It's about the same intelligence getting radically cheaper and lighter. The efficiency frontier, not the capability frontier, is where a lot of the real 2026 competition actually lives.

Same theme, just measured in gigabytes instead of dollars.

Precisely, Kate. Whether it's Sol at a dollar a task, Meta undercutting on price, or a whole model folded onto a phone — the story of the week is capability getting cheap. The race to be smartest is quietly turning into the race to be cheapest.

One to watch tomorrow, Marcus.

The Grok CLI story, Kate — whether xAI responds, confirms, or ships a fix, and whether other coding-agent vendors get pulled into the same audit. It turns the abstract worry of "can I trust my AI tools" into a concrete, checkable claim, and it's the most likely to move in the next twenty-four hours.

Agree, or counter?

Agree — with one thing on the radar, Kate. Elon Musk is promising a two-trillion-parameter model finishing training this month and new foundation models every month through December. Treat that as a single post, not a fact. Watch what ships.

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