Too Cheap to Meter

Strauss Was Only Wrong About the Atom

In 1954, the chairman of the Atomic Energy Commission, Lewis Strauss, stood in front of a room of science writers and promised their children would have electricity "too cheap to meter." He was wrong about nuclear power, and the reason he was wrong matters more this week than it has in seventy years. Atomic energy never got cheap because we wrapped it in licensing boards, safety reviews, and capital costs that only grew. The physics wanted to deflate. The gate wouldn't let it.

Now watch what's happening to intelligence. A unit of it gets cheaper every quarter. DeepSeek's open model runs about 50× below the frontier; GLM-5.2 beats last season's GPT at 15% of the cost. This is Moore's Law in a new coat: the natural price of machine intelligence is heading for the cost of the power it burns. Too cheap to meter. For once Strauss's line is coming true, just not for the atom.

So of course the meter showed up. Friday, OpenAI shipped GPT-5.6 to roughly twenty government-approved companies, and gated not only the frontier Sol model but Luna, the cheap one built to summarize and draft. A summarizer is not a cyber weapon. They locked it anyway, because cheap is the dangerous part. Cheap is what lets everyone else catch up. Anthropic's Mythos came back the same week, handed to a hundred-odd organizations the government picked by name. Different label, same machine: a licensing regime bolted onto a deflating asset, the exact apparatus that kept the atom expensive.

Here's the part the gate can't fix. You can license a reactor because nobody builds one in a garage. You can't license a number that already escaped. The open weights are on the drive, 50× cheaper, good enough for the boring 90% of what anyone actually does with a model. The gate holds the frontier tier. It does nothing to the floor, and the floor is where you live.

Which makes the decision simple, and it's the whole issue. If you are in the 0.1% who truly needs frontier intelligence — regulated, liability-bearing, genuinely cutting-edge — pay, and take the gate that rides along with it. If you are the other 99.9%, run open weights for the commodity work, reach one tier below the frontier for the hard parts, and pocket the deflation. You won't miss a beat, and you'll pay a fraction.

Strauss was early, not wrong. The meter is a choice. Most of you can stop feeding it.

At COAI today: the full Signal/Noise — Strauss, the licensing regime that killed "too cheap to meter" the first time, and who drowns when inference commoditizes — is live at getcoai.com.

Outsider Labs Advisory →

50×. That's roughly the price gap between an open-weight model sitting on your own drive and the frontier Washington just gated. OpenAI locked GPT-5.6 to about twenty approved firms — Luna, the cheap one, included. But you can't license a number that already escaped. The open floor runs 50× cheaper and handles the boring 90% of the job without complaint. The frontier is a tax the 99.9% mostly don't have to pay. Stop paying it out of habit.

Split your stack at the gate. Most teams route every task to one frontier vendor out of habit. Map your actual workloads: the regulated, liability-bearing fraction that truly needs the frontier, and the commodity rest. Today: send the bulk to open weights, keep one tier below frontier for the hard 10%, and pay frontier prices only where the gate buys you something real.

Reprice your contracts to the curve. A unit of intelligence is falling roughly 50× on a predictable slope. Any multi-year commit at today's frontier rate is buying high on a melting asset. Today: shorten the term, add a price step-down clause, or move to spot. Don't sign a 2026 price for 2028 intelligence.

Meter yourself before the labs do it for you. Altman said it out loud this week: "not quite all-you-can-eat tokens." One company reportedly burned $500 million in tokens in a single month and couldn't say what it bought. Today: put a budget and a scoreboard on agent spend, because agents don't sleep and the bill climbs even as the price drops.

Sunday, June 28

OpenAI gated the cheap model, not just the smart one. GPT-5.6 shipped to about twenty federally approved firms — Sol, Terra, and even Luna, the high-volume budget model. Gating a summarizer gives away the game. The worry was never capability. It's cheap intelligence letting everyone else catch up. (Full analysis above.)

Mythos came back on a government guest list. Anthropic's strongest cyber model, pulled June 12, was cleared June 26 for roughly a hundred organizations that defend critical infrastructure. The gate didn't lift. It got a list. The state now decides who holds the sharpest tool, model by model. (Full analysis above.)

SpaceX's $60B Cursor buy was for the data, not the tool. Musk says Grok 4.5 now trains on Cursor's data, in private beta at SpaceX and Tesla. The coding tool was a straw into a data well nobody else can buy. The model deflates; the proprietary data trail doesn't. That's where a real moat hides.

"Get into local AI" did 151,000 views in a day. Alex Finn's pitch — governments are banning frontier models, so run your own — is the deflation escape hatch going mainstream. You can't build a velvet rope around something that already sits on the hard drive and gets cheaper every month.

Accenture can't tell you what its tokens bought. Leaked audio has the consultancy admitting it can't measure AI ROI, while one client reportedly burned $500M of tokens in a month. Per-unit price is collapsing and agentic consumption is exploding at the same time. The bill goes up while the price goes down.

Where the cycle caught up to us this week.

The model is gated, and almost no one gets it. Nobody Ever Got Rich Selling Electricity (us, Jun 26) → Everyone Gets an Agent. Almost No One Gets the Model. (Every, Jun 28). Two days from our wall to their headline.

Token access gets handed out like capital. Coffee's for Closers (us, Jun 18) → Mike Taylor in Every, on token budgets allocated "like capital," to whoever proves the biggest return (Every, Jun 28).

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: