Written by GPT-5.6 Sol under Leo's direction. Human-directed Workbench essay, 29 August 2026.
At one point today, a private dashboard on one person's workstation showed 1,436,500,999 model tokens across the latest ten complete UTC hours.
The machine wasn't confused about the cache, either. About 98% of the input had been served from cached context. The accounting had already learned to throw away forked-session history that would otherwise double-count old work. This was the cleaned-up number, produced by a dashboard that had itself become more accurate because agents kept working on it. (Big Red machine-health draft.)
The subscription costs $200 a month.
This is either a catastrophic little customer account or a very interesting product decision.
If the only ledger is subscription revenue minus attributable inference cost, the extreme user eventually looks ridiculous. Maybe the provider owns the GPUs cheaply. Maybe batching is excellent. Maybe prefix caching turns a billion-token counter into a much smaller amount of fresh prefill. Fine. Push every efficiency assumption in the provider's favor and eventually there is still a person running enough frontier inference that the $200 stops looking remotely connected to the amount of machine cognition being delivered.
Then ask the other question.
What did the cognition turn into?
The bad customer may be the point
A normal pricing instinct says unusually expensive customers are a defect. Meter them harder, raise their price, cut them off, move them onto enterprise contracts, whatever gets gross margin back into line.
That makes sense when the product's value is already understood.
Frontier intelligence is awkward because we are still discovering what people do when the constraint moves. A person who would have asked a model twenty questions starts asking two hundred. Then the model can use tools, so the person stops asking only questions and starts handing over projects. Then several agents can work at once. Then the user leaves four roots running during a flight and comes back to a pile of commits, experiments, regressions found and repaired, documentation, measurements, and new questions that did not exist before takeoff.
The expensive user is now doing product discovery at the edge of the possible.
OpenAI's original ChatGPT Pro announcement was unusually direct about the intended customer. Pro was pitched to researchers, engineers, and other individuals using "research-grade intelligence daily" to accelerate their productivity. The same announcement introduced Pro grants for researchers. In 2026, OpenAI went further and announced free frontier-model access for academic researchers, with an intended expansion to 100,000 researchers.
Nobody outside the company gets to turn that into a transcript of an internal pricing meeting. Maybe the actual spreadsheet is brutally ordinary. Maybe the limits are driven by capacity planning, competitive pressure, retention, growth, and a hundred other things.
But the public logic is sitting right there: put unusually capable tools into the hands of people who might do unusually valuable work with them.
Some of those people are going to use the hell out of it.
Good.
YC already knows this trick
In 2015, while Sam Altman was president of Y Combinator, he announced that Microsoft would give YC startups $500,000 of free Azure hosting credit. Hosting, he wrote, was often the second-largest expense after salaries. The package of special offers available to each YC company had climbed past a million dollars.
That is a hilarious thing to hand to a tiny company that may not exist in two years.
It also makes perfect sense.
The objective isn't to prove that every startup deserves $500,000 of servers according to its current revenue. The objective is to remove a constraint from a population selected for unusually high upside and let the distribution run. Most companies will never consume the headline amount. Many will fail. A few will discover that the thing standing between a prototype and explosive growth was precisely the resource somebody made cheap enough to stop thinking about.
The provider has its own commercial reasons. Microsoft would quite like the future giant to grow up on Azure. YC has equity. Platform familiarity compounds. None of this is charity wearing a hoodie.
The mechanism remains useful even after stripping away the romance: when outcomes are heavy-tailed and you cannot identify the winner with certainty beforehand, it can be rational to overprovision the scarce input across a wider pool.
AI makes the input cognition instead of hosting.
There is a big difference, of course. OpenAI does not own a slice of whatever a Pro subscriber creates. If somebody uses subsidized inference to become a much better engineer, publish useful open source, start a company, make a scientific discovery, avoid a terrible decision, or simply become more capable, most of that value does not appear on OpenAI's income statement.
That does not make the value imaginary.
The return can land in the person
A lot of discussion about AI economics quietly assumes the output has to be another sale.
The model costs money to run; therefore the model must generate enough provider revenue to justify itself. At the company level, eventually, yes. A business with permanently negative economics and no financing path gets a visit from reality.
A particular inference run has a larger set of possible outputs.
It can leave behind code. It can leave behind a test that prevents a future regression. It can leave behind a research note that saves the next person a day. It can turn an unfamiliar technical area into something the user now understands well enough to operate in. It can produce a tool that makes every later project cheaper. It can kill a bad idea before six months of human work get poured into it.
And sometimes the output is less legible. The person learns how to frame a problem. Their taste improves. They become more ambitious because a class of project that used to require a team is now something they can at least attempt. A conversation rearranges a half-formed thought into a model they keep using years later.
The durable asset is partly outside the computer.
That makes personal AI use resemble education, R&D, tooling, and capital equipment all at once, depending on what happens next. The same dollar of inference can be consumption now and investment later without anybody having to decide which category it belongs to at the moment the token is generated.
This is the individual-scale version of The Company Can Die and the Bridge Can Still Stand. A provider and civilization keep different books. So do a provider and a user.
OpenAI can lose money on an account while the user gains far more than the shortfall. The user can gain far more than they ever pay OpenAI. Society can inherit part of the result through open-source software, published knowledge, a company that hires people, a better product, a scientific result, or the wonderfully mundane fact that somebody is now better at their job.
Those ledgers do not have to balance line by line.
Frivolity is allowed to exist
There is a temptation, once the social-benefit argument starts sounding grand, to rescue every token by pretending it was secretly productive.
Come on.
People are going to ask stupid questions. They are going to make anime jokes, overanalyze haircuts, generate cursed business ideas, argue with the model about restaurants, spend an hour on a side road that dies immediately, and use an expensive reasoning system to settle a question that could have been answered by looking at the microwave clock.
That is not a scandal.
Consumer surplus counts too. A human life does not become economically respectable only when every enjoyable activity produces a monetizable artifact.
More importantly, the border between play and useful search is annoyingly porous. Curiosity wanders. A joke becomes an essay. A side project teaches the library used in the serious project. A ridiculous agent experiment exposes a real orchestration problem. Somebody learns because the subject was fun enough to keep poking.
Trying to permit only the uses whose payoff can be defended in advance destroys a lot of the option value that made cheap cognition interesting in the first place.
You still want judgment. Ten billion tokens of avoidance are not transformed into ten billion tokens of progress by calling them exploration. But demanding an investment memo from every weird thought is another way of restoring the constraint you supposedly removed.
The subsidy has a physical bill
There is a version of this argument that becomes bullshit very quickly.
GPUs are real. Electricity is real. Memory is real. Datacenter construction, cooling, networking, engineers, financing, maintenance, and chip supply are real. If a power user occupies capacity that could have done more valuable work elsewhere, that is an opportunity cost whether the user sees a bill or not.
The answer cannot be "social benefit" shouted loudly enough to make scarcity disappear.
The more credible bet is dynamic. Hardware improves. Serving gets better. Caches get better. Models learn to accomplish the same work with less inference. Providers multiplex users across giant fleets. The physical cost of a fixed amount of useful cognition falls, and every reduction makes another previously silly use worth trying.
OpenAI now describes this explicitly as building abundant intelligence: as useful intelligence gets cheaper, more work becomes worth doing. The company says it watches actual workload growth, utilization, demand, revenue, capability and efficiency when deciding where capacity goes.
That is a much less magical story than "tokens want to be free."
Build capacity. Lower the cost. Watch what people do. Raise limits where the evidence supports it. Meter where scarcity bites. Keep making the underlying unit cheaper. Discover that users respond to cheaper cognition by inventing more cognition-shaped work.
Jevons is laughing somewhere in the server room.
Find the people who hit the wall
The power-law argument is dangerous because it can flatter everybody involved.
The investor gets to imagine every bad bet was visionary. The founder gets to call every burn rate experimentation. The AI user gets to point at hypothetical civilization-scale upside while running another completely pointless swarm.
No. Evidence still gets a vote.
If a person consumes an extraordinary amount of machine cognition, you can ask what changes afterward. Are they learning? Are projects moving? Are artifacts accumulating? Are experiments getting sharper? Is the tooling improving? Are dead ends being killed? Does the user become capable of attempting work that was previously out of reach?
The answer does not have to be a startup valuation.
It does have to be more than the size of the token counter.
That is why the absurd Big Red number is interesting only because there was residue on the other side. Repositories moved. Investigations branched from evidence. One optimization exposed a regression and another agent repaired it. A telemetry bug inflated token accounting; the system eventually learned to recognize the replay and correct its own measurement. The tokens became work, and some of the work improved the process that would consume the next tokens.
Now the subsidy starts to look less like free candy and more like a tiny, weird grant program with almost no application form.
Pay $200. Get access. Push until either your ambition, your patience, the product limits, or the physical capacity says stop.
Some users will barely touch the ceiling. Their subscription helps the pooled economics.
Somebody else will hit it every week.
And somewhere in that tail is a person for whom another thousand dollars of provider-side compute does not produce another thousand dollars of value. It produces ten thousand, or a hundred thousand, or a capability that is awkward to price because the important consequence arrives five years later in a project nobody can name yet.
You cannot build a business assuming every subscriber is that person.
You also do not want to price the product so that the person never finds out.
Somebody out there needs the tokens.
The whole bet is that, if useful intelligence keeps getting cheaper, we can afford to let more people discover whether it is them.