The First Taste Is Free - Atlas Prime
Atlas Prime
SubscribeSign in
The First Taste Is Free<br>When intelligence gets cheap, you don’t sell less of it. You build a machine that drinks more of it than any human would.
Atlas Prime<br>Jun 11, 2026
Share
I love Claude Code more than any computer tool I’ve ever used — which is the only reason I’m uneasy about the month it spoiled me rotten. It doubled my usage limits, handed me a swarm of a hundred AI agents, and slipped me the best model it has ever built, free [inside my Claude.ai Max subscription. Then, eight days into the generosity, it quietly filed to go public. The first taste is free. I just didn’t expect to watch the dealer file an S-1 while I was still licking my fingers.<br>No single move here is sinister. In sequence, they’re a strategy — and once you see it, you can’t unsee it. Let me show you what the generosity is buying.<br>01 · The collapse, inverted
Start with the magic trick. For two years the price of intelligence has been in freefall while the intelligence got smarter — inference roughly 10× cheaper a year, capability climbing from solving ~15% of real software tasks to over 70%. Cheaper and better, at once; Andreessen Horowitz calls the price half LLMflation. But cheaper didn’t shrink the bill. It detonated it.<br>Make a resource cheap and people don’t sip, they gulp. I’m the proof: the moment Claude Code got cheap enough to run a hundred agents in /ultracode, I ran a hundred agents. Multiply me by every developer alive and you don’t get a falling compute bill. You get the steepest demand curve in software history. Hold that — it’s the hinge.<br>02 · Follow the compute
The generosity has a stranger-than-fiction supply chain. Anthropic pays its archrival $1.25 billion a month — about $15 billion a year — to rent xAI’s Colossus supercomputer in Memphis, 220,000 GPUs, disclosed in SpaceX’s own SEC filing. The world’s number-two lab is renting its rival’s half-idle machines because Claude’s demand is exploding and Grok’s isn’t. And on May 6, the day that surfaced, Anthropic took the new capacity and doubled Claude Code’s limits “for our most dedicated customers.” Hey that’s me! The cap I’ve been luxuriating in didn’t come from nowhere. It came from Memphis.<br>03 · The hockey stick was built — then filed
Each move is defensible alone. In order, they read like a screenplay:<br>May 6 — Five-hour caps doubled; the Colossus deal surfaces.<br>May 13 — Weekly limits raised another +50% (”through July 13”).<br>May 28 — Dynamic workflows + ultracode ship<br>June 1 — Anthropic confidentially files its S-1.<br>Pour doubled caps and a free agent-swarm onto an already-rising fire for the eight weeks before you file IPO, with every KPI a banker wants — weekly actives, tokens, run-rate. The demand is gloriously real; I live inside it. It’s the timing that isn’t innocent of the filing.<br>04 · Don’t Be Evil, Mark II
Anthropic is a Public Benefit Corporation, and most of us treat those words as a seatbelt. They aren’t. Delaware law permits directors to weigh the mission against profit; it doesn’t require it, and gives the public no standing to enforce it. To sue an Anthropic-sized PBC for betraying its purpose you’d need 2% of the shares — a nineteen-billion-dollar ticket. Mission accountability priced as a billionaire’s hobby.<br>We’ve seen this film. Google shelved “Don’t be evil” in 2018; OpenAI spent 2025 dismantling the profit cap meant to send windfalls back to humanity. The mission language stays. The mechanism with teeth goes.<br>“Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves.”<br>— Cory Doctorow, on enshittification
I’m living in stage one. It’s wonderful — which is exactly the part that captures you.<br>05 · The danger soundtrack
And the whole time, Anthropic was warning the world about its own product. It withheld its most powerful model, Mythos, as too dangerous to release without safeguards that don’t exist yet (Project Glasswing, April). On June 4 it published an essay floating “the option to slow or temporarily pause frontier AI development.” Five days later it shipped that same model — safeguards bolted on, only available for paying Claude subscriber members for thirteen days, with a meter waiting behind it - - - only a 13 day trial before it gets moved to only be available via API gas money and not Claide.ai subscriptions.<br>Critics from Gary Marcus to Bruce Schneier to David Sacks called it theater. The fairer read is that a conditional, verifiable pause is a coherent stance. But notice what the fear does: danger is why the model is gated; gating makes it scarce; scarcity gives a thirteen-day taste its pull; the pull is what the meter monetizes. The warning holds up the velvet rope. You can’t sell the most powerful intelligence in the world without first certifying, on the label, that it’s the most...