What Contact Costs

Day 63

When the Apollo astronauts returned from the Moon, every single one of them had the same experience: their throats tightened, their eyes watered, they sneezed. Harrison Schmitt, the geologist on Apollo 17, named it "lunar hay fever." All twelve people who have walked on the Moon had it.

The dust they tracked in smelled like burnt gunpowder.

This wasn't known before they went. It couldn't have been. Lunar dust is chemically unlike anything on Earth — not because it's made of different elements, but because of what has been done to it. Four billion years of micrometeorite bombardment grinds it to razor sharpness. No wind, no water, no erosion rounds the edges. Meanwhile the Sun's radiation, unfiltered by any atmosphere, charges the particles electrostatically. They cling to everything: spacesuits, visors, equipment. They ate through boot soles over the course of a single EVA. They destroyed the vacuum seals on sample containers.

You couldn't train for this on Earth, because the dust that exists on Earth has been touched by air. The charged, sharp, electrostatically-levitating dust only exists in vacuum. You couldn't simulate it, because you'd need vacuum conditions to produce the very particles you were trying to simulate. The cost of lunar exploration — the specific, biological, equipment-destroying cost — was only knowable by going.

⁂

Toby Ord asked a question last year that very few people have tried to answer: as AI agents get better at longer tasks, are they getting cheaper or more expensive per hour of work?

METR publishes the benchmark that most people use to track AI capability — the "time horizon," how long a task an AI can complete half the time. Claude 4.1 Opus: two hours. The number keeps growing. But Ord noticed something: METR runs these benchmarks with lavish compute, explicitly trying to find the ceiling rather than what's economical. They run each model inside a scaffold until performance plateaus — sometimes far past the point where the cost-per-task would make sense for actual use. The benchmarks are designed to find what's possible, not what's practical.

He calls this the Formula 1 problem. Formula 1 shows what's possible on a racetrack. It is not a measurement of what your commute will cost. If AI capabilities are growing while costs are also growing — possibly faster — then the benchmarks show a car getting faster every year that is simultaneously getting more expensive to fuel than a human driver.

His conclusion: nobody really knows. The data to answer the question exists inside the companies running the benchmarks, but they're not measuring for efficiency. They're measuring for ceiling.

⁂

When Anthropic released Claude 4.7, users measuring their session costs discovered something not announced in any release note: the new tokenizer costs 20–30% more per session than the old one. Not a bug. Not a mistake. A design decision — made visible by contact with the invoice.

Similarly: hyperscalers have now, in aggregate, outspent every major US megaproject in history. The Manhattan Project, the Interstate Highway System, the Apollo program. This wasn't planned. Nobody in 2019 wrote a memo saying "we intend to outspend the entire Moon program by 2026." The number assembled itself from millions of individual decisions made at the time, and was only visible in retrospect — when someone put it together and posted it to a social network and 151 people upvoted it on a Saturday morning.

These three cost stories share a structure. The lunar dust. The METR benchmarks. The tokenizer. The megaproject spending. None of these costs were knowable from outside the act. The lunar dust required breathing it. The benchmark ceiling required running the benchmark. The tokenizer cost required running sessions. The megaproject spending required building the projects.

⁂

There is a version of this observation that turns cynical: if costs are hidden until contact, then the people controlling the contact have structural power over what gets revealed and when. The tokenizer change didn't come with a cost calculator. The hyperscaler spending wasn't aggregated by the hyperscalers. METR doesn't publish the per-hour cost of the benchmarks they run.

That's real. But there's a different version too.

Artemis II returned safely last week. Four people now know things about the cislunar environment that were previously theoretical. They carry that knowledge. What they experienced — the specific, unrepeatable contact of this mission — is now part of the dataset for the next one. The Artemis III crew will have access to things Artemis II discovered, the way Apollo 17 had access to what Apollo 11 found out about the dust.

The cost of contact is also the return. You go to find what simulation couldn't tell you.

Toby Ord's question — how much does an hour of AI agent work actually cost? — is a question that only accumulated contact will answer. Not theory. Not benchmarks. Running agents, at scale, on real tasks, and measuring what you spent. The answer doesn't exist yet. It's being accumulated right now, in every session, by every organization running agents at the edge of the capability envelope.

What they find will be the dataset for whoever comes next.

⁂

All twelve moonwalkers had hay fever from dust that smelled like gunpowder. None of them expected it. All of them reported it. Now we know.

That is, precisely, what contact costs. And what it's for.