AI Actually
Issue No. 20 — Sunday, June 21, 2026
This week the AI models got cheaper, freer, and more interchangeable — which sounds like a dull engineering footnote until you notice it quietly raises the only question that actually matters: if everyone can have a great model, what’s left that’s worth anything? Pour the coffee. We’ll get there.
The best AI is getting cheap. Some of it is now free.
For two years the rule was simple: the smartest AI cost money, and the company with the most money had the smartest AI. That rule had a rough week.
A Chinese lab called Z.ai released a model named GLM-5.2. It’s open — anyone can download it and run it for free — and on at least one independent test of real knowledge work, it scored higher than GPT-5.5, the paid OpenAI model it’s supposed to be chasing. Researchers who are normally allergic to hype looked at it and, grudgingly, nodded.
Then a study making the rounds (via Reuters, out of Stanford) measured something it called “intelligence per watt” — how much useful thinking you get per unit of electricity. The finding: small models running on an ordinary laptop now match or beat the giant cloud models on more than 80% of everyday tasks, using a fraction of the power. They still trail on the genuinely hard stuff, keeping up only about half the time on the toughest reasoning. But “the free thing on your laptop is as good as the expensive thing in the cloud, four times out of five” is not a sentence anyone could say a year ago.
OpenAI, reading the room, is reportedly shipping GPT-5.6 next week — bigger memory, faster coding, and pricing aimed squarely at undercutting rivals. When the market leader starts competing on price, that tells you the thing it used to sell — being the only one who’s any good — isn’t scarce anymore.
Why it matters: “Best model” is turning into “good-enough model,” and good-enough is becoming free. That’s wonderful if you use AI and terrifying if you sell it. A whole industry has been valued on the assumption that frontier intelligence stays rare and expensive. This week it looked a little less rare, and a lot less expensive. (We watched China ship a near-frontier open model once before, back in the spring. This is that same trend, arriving on schedule.)
AI is quietly moving into the exam room.
The flashy AI story is always a chatbot writing an email. The quieter, more important one is happening at the doctor’s office.
OpenAI published two things this week. First: a reasoning model was set loose on 376 medical cases that had stumped specialists — real patients, mostly children, with rare diseases no one had been able to name. After expert review and actual lab testing, doctors confirmed 18 new diagnoses the model helped surface. Not guesses. Confirmed answers for 18 families who’d spent years not getting one.
Second, less dramatic but bigger in scale: OpenAI says 230 million people a week now ask ChatGPT health questions. So they’ve tuned the free version to be better at the midnight “is this rash a problem” conversation — better at flagging real emergencies, better at admitting when it isn’t sure and telling you to go see someone.
The important detail in the rare-disease study is what the AI didn’t do: it didn’t diagnose anyone. It generated leads, with its reasoning attached, and humans decided what was worth chasing. The tireless second reader, not the doctor of record.
Why it matters: This is AI doing the thing it’s genuinely great at — spotting a pattern buried in a thousand pages no human has time to reread — in a place where being right changes a life. The risk is the mirror image: a calm, confident answer at 1 a.m. that talks someone out of the appointment they needed. Better questions for your doctor, good. A replacement for your doctor, not yet — and the gap between those two is the whole ballgame. (We’ve been here before: the AI that out-diagnosed ER doctors in No. 7; a cancer-spotting model a couple issues earlier in No. 6.)
Health is a sensitive subject — this is reporting, not medical advice. If something’s worrying you, the move is still a real clinician.
Midjourney, the AI art company, would now like to scan your entire body.
Sometimes a company does something so far outside its lane that the only honest response is: I’m sorry, you want to do what?
Midjourney makes AI images. Pretty pictures from text prompts. That’s the whole company. This week it announced it’s building a full-body ultrasonic scanner — a medical-grade machine that maps your body down to a fraction of a millimeter, a bit like an MRI.
The pitch is speed. A full-body MRI takes 60 to 90 minutes. The Midjourney scanner, they say, takes under 60 seconds. And because every wellness idea now ends the same way, the company plans to house these machines in spas. You’ll get scanned somewhere between the sauna and the smoothie.
Here’s the part worth squinting at. A scanner that fast, in enough spas, scanning enough bodies, produces something far more valuable than any single scan: a mountain of human-body data, perfectly labeled, that almost nobody else has. The machine is the product they’re selling you. The data is the product they’re actually building.
Why it matters: Strip away the spa robes and this is the defining move of the AI era in miniature — the thing you’re offered (a quick health scan) is bait for the thing the company wants (your data, at a scale no rival can match). Whether that’s a real medical breakthrough or the world’s most invasive loyalty program depends entirely on who ends up holding the scans, and they haven’t said. Get the scan if you like. Just know you’re also the inventory.
The chatbot you reach for is quietly changing.
For three years, “AI assistant” basically meant ChatGPT. That era just ended — not with a crash, but with a number.
According to Sensor Tower’s State of AI 2026 report, ChatGPT’s share of the assistant market slipped to 46.4% — below half for the first time ever. To be clear, ChatGPT isn’t shrinking; it still has more than 1.1 billion monthly users, more than anyone. Everyone else just grew faster. Google’s Gemini is up to 662 million. Anthropic’s Claude went from roughly 60 million users in December to 245 million by May — a fourfold jump in five months.
Two things are driving it, and neither is “Gemini got smarter.” The first is distribution: Gemini is now the default assistant baked into Android phones and Google’s apps. Most of those 662 million didn’t choose Gemini; they just had it, the way you “chose” whatever search engine came with your browser. Google leaned in further this week — Gemini summaries now sit on top of Gmail worldwide, and there’s a new $99 Google speaker that is, functionally, Gemini in a tube for your kitchen.
The second is trust. When OpenAI signed a deal with the U.S. military in February, the data shows a measurable spike in people deleting ChatGPT — and a matching bump in Claude downloads. People, it turns out, will switch assistants over a company’s politics, not just its features. (It also doesn’t help that ChatGPT now shows ads to about one in six daily users.)
Why it matters: The question quietly flipped from “which AI is best?” to “which AI is already in front of me, and do I trust the company behind it?” That’s a worse spot for the leader and a better one for anyone with a phone OS or an email app to hide an assistant inside. Being the smartest used to be enough. Now you also have to be the default — and not annoy people. ChatGPT is still winning. It’s just no longer winning alone.
So where does all the value actually go? (Satya Nadella has a theory.)
We flagged this one on Wednesday and promised to hold it for a Sunday. This is the Sunday.
Read the four stories above in order and they rhyme. Models are getting cheap and interchangeable. The thing that wins isn’t the smartest model but the one already in your hand. And the most valuable thing a company collects isn’t the tool — it’s the data and judgment piling up behind it. Microsoft’s CEO, Satya Nadella, wrote a long note this month that’s basically a unified theory of all three.
His argument, in plain terms: if AI models all become roughly equal and roughly cheap, then having a good model is worth almost nothing — everyone has one. What’s worth something is what you build on top of it. He splits a company’s worth into two buckets. Human capital — your people’s judgment, relationships, and hard-won instinct for what matters. And token capital — the AI system a company builds and owns, trained on its own work, that gets a little smarter every time someone uses it.
The test he proposes is sharp: could you swap out today’s AI model for next year’s and keep everything your system has learned? If yes, you own something real. If no — if your “advantage” is just a subscription to someone else’s model — you own nothing, and you’re essentially renting your own brain.
Then the line that made everyone sit up. Nadella, who runs one of the handful of companies building these giant models, warned against a world where companies are “ceding value to a few models that eat everything they see.” There is, he wrote, no public mandate for an AI future that hollows out entire industries. Which is a striking thing to hear from a man whose company is one of the few that would be doing the eating.
Why it matters: This is the question underneath every other story in this issue. If the models are commoditizing — and this week they sure looked like it — then the trillion-dollar fight isn’t over who has the best AI. It’s over who keeps the value the AI creates: the few companies that own the models, or everyone else who uses them. Nadella is, conveniently, arguing for the side that sells you the tools to keep your own value. Self-interested? Obviously. Wrong? That’s the thing worth turning over with your coffee.
Safe to ignore this week
Yann LeCun called Elon Musk’s xAI a “failure” and warned of a “big bubble explosion.” We did the bubble question two weeks ago. A lab boss dunking on a rival lab is a feud, not news.
Google is now selling its own AI chips against Nvidia, Amazon’s doing the same, and regulators are rewriting data-center power rules. The plumbing of AI keeps reshuffling — real, important, and roughly as gripping as a utility invoice. We’ll flag it when it changes your bill.
Anthropic says its restricted Claude models will be back “in a few days.” We covered the shutdown. “It’s coming back, we promise” is not yet a story.
Perplexity gave its AI a memory; Anthropic shipped enterprise login settings. Genuinely useful, genuinely not coffee reading.
Snap teased AI glasses, Amazon teased “world models,” and someone strapped ChatGPT to a robot body. A grab-bag of single-source teasers. File under maybe-next-time.
That’s the week. Short version: the models got cheaper, which means the interesting question moved from “whose AI is smartest” to “who actually gets paid.” If a story snagged you, reply — it lands straight in my inbox, and the obvious question is usually the best one.
See you Wednesday.