The autonomy ceiling is a spending decision

Earlier this summer, an unreleased Anthropic model worked on a single math problem for about 36 hours without a person driving it. It spun up 60 copies of itself to check each other’s arithmetic, search the literature, and rebuild proofs independently. It pushed a 165-year-old open question, the Riemann hypothesis, further in one run than the field had managed before. And before that run worked, it failed 650 times.

That run, and a few like it, are the reason a lot of people have quietly concluded that agent autonomy has no ceiling worth planning around. If a model can grind on a research problem for a day and a half and come back with something real, then today’s short leash is temporary, and any roadmap built around agents doing longer and longer stretches of unsupervised work is just betting on the trend.

The trend is real and measured. METR, a research group that clocks how long AI agents can work unattended before they fail, puts the reliable production ceiling at two to four hours. More important than the number is the slope: that ceiling has roughly doubled every seven months for six years. Extend the line and the multi-day research run stops looking like a stunt and starts looking like next year.

The line leaves out where the doubling comes from. It happens because someone keeps buying longer context windows, bigger reasoning budgets, and more coordinated subagents, and shipping them. It’s a purchase, renewed every model generation. And the company doing most of the buying just filed to go public.

The ceiling isn’t a wall the models are approaching. It’s a spending decision, and that decision is about to get a public-market referee.

Walk through what changed. Anthropic filed its confidential S-1, the paperwork a company files to sell shares to the public, on June 1, and is targeting an October listing. Inside that filing is a number no outside party has ever seen: gross margin, the share of revenue left after paying the cloud bill to run the models. Anthropic’s projection has it climbing from around 50% this year to 77% by 2028, on the back of roughly $80 billion in cumulative cloud infrastructure through 2029 [both reported]. Once the company is public, that margin gets reported every quarter, to shareholders who will price the stock on it.

And the cheapest way to move a margin number in the right direction is to stop paying for your most compute-hungry features. Long context. Multi-step reasoning that runs for hours. Dozens of agents coordinating on one task. In the monitor’s own read of the filing, those are exactly the features a margin-disciplined public company is structurally pushed to deprioritize, in favor of deployment density and cost-per-inference [reported]. That list is not random. It is a precise description of the Riemann run.

So the evidence everyone points to as proof the ceiling can be beaten is also the single most margin-exposed thing in the building. Anthropic’s other headline demo, a Mythos Preview model that ran “largely autonomously for several days” to find novel cryptographic weaknesses, cost roughly $100,000 in API usage per discovery [reported]. The Riemann run burned 31 million output tokens over those 36 hours [verified]. The token bill for the winning run alone was cheap, around $310 [verified]. The real cost was the 650 failed approaches that came first. What prices a sustained research program is the capital it takes to survive the runs that produce nothing. The win is a rounding error against the losses that paid for it.

That is the definition of a loss leader. It’s the kind of capability a lab runs once, publishes, and puts in a filing to prove it sits at the frontier. It is not the kind of capability it necessarily sells you as a product tier and sustains at scale. The capability-ceiling case has been counting the demo as if it were the trend.

The honest counterargument is that the doubling looks like a law, not a choice. It held across six years and several model generations, through different companies, and it shows up in released capability whether or not any single lab wants it to. On that reading, one company’s IPO is a footnote about its feature cadence, not a dent in the trajectory. That’s a fair reading, and it’s the one to beat.

Two things bend it. First, ask who actually carried the curve. If the six-year doubling was driven disproportionately by the specific frontier labs now walking into public markets, then it’s not obviously independent of their spending, and a margin referee is a real variable, not a footnote. Second, look under the ceiling. Production agent systems already route 70 to 80% of their model calls to small, cheap models running locally, because most tasks never needed the frontier [reported]. The money already flows to the plentiful work below the ceiling. Margin discipline pushes the same direction it’s already going, and away from the expensive 20% that lifts the ceiling at all.

That splits the question in two. “How high can agents go” and “how high is it margin-positive to let agents go” are not the same question, and they only look identical while one company is willing to eat the cost of the gap. The Riemann model could almost certainly go further. The open question is whether, after October, its maker keeps choosing to pay for the further.

There’s a clean way to watch this resolve. Anthropic’s Q3 gross margin lands in October, the first time that number faces public shareholders. In the months after, watch whether the release cadence splits. Track how often the compute-heavy features ship next: longer context, higher reasoning-budget ceilings, bigger multi-agent limits. Then track how often the cheap ones ship: smaller models, better routing, lower cost-per-inference. If those two cadences pull apart after October, you’re watching a margin target edit a product roadmap in real time.

For a PM, this is a horizon assumption hiding inside a roadmap. Betting that agents will routinely cross today’s few-hour ceiling in the next 18 months isn’t a bet on a capability curve. It’s a bet that the company carrying that curve keeps paying for it once the bill is read aloud every quarter. That might hold. Harrison Rolfes at PitchBook, one of the analysts watching the filing, has said the margin number “will either validate or collapse the entire narrative the private markets have been pricing for three years” [reported]. The trajectory you’re planning around is a line on that same statement. October is when we find out whether it’s a law or a purchase order.