Toby Ord argues that giving an AI model ten times as many tokens does not buy what giving a person ten times as much time buys, and that newer models appear to be narrowing the difference. He made the case in an essay on his own site dated October 9, 2026, and describes his evidence as “suggestive” rather than settled.
A token is the small chunk of text a model reads or writes. The number a model may spend on a task works as its thinking budget: long chains of reasoning, repeated tool calls, retries. Many people treat that budget as the machine equivalent of giving a worker more hours. Ord asks whether the exchange rate holds. If a person improves a result by a given amount with ten times the time, does a model need only ten times the tokens to match that, or a hundred, or ten thousand?
He says the answer matters because the analogy shapes how people reason about AI. If models need a hundred times the tokens to match a person’s tenfold gain from more time, he writes, the comparison “has broken down”. It could also be that the comparison is fine and models are simply worse at using extra time, a weakness he says should be tracked. And they could end up better at it than people.
The first data he cites comes from METR, which charted model and human scores on research-engineering tasks at several budgets. Ord reads it this way: each time the budget doubles, a person’s score climbs about four times further than a model’s does. Carried forward, that would leave a model needing roughly 10,000 times the resources to match what a person gets from ten times the time. He then lists the chart’s limits himself: the models were early ones, only o1-preview was a reasoning model, and no human worked longer than eight hours.
His better evidence is a chart the UK AI Security Institute released in July 2026. It plots how far a model’s “80 percent time horizon” stretches as it gets more tokens. That horizon is the length of task, measured by how long a human takes, that the model completes 80 percent of the time. The tasks were the institute’s own capture-the-flag hacking challenges.
Ord’s reading is that every model’s line climbs more slowly than its token budget, so ten times the tokens always buys less than ten times the horizon. By his estimate, the earliest model in the set needed around 10,000 times the tokens for a tenfold longer horizon. The most recent, covering models up to May 2026, needed about 20 times. He says he used Claude Opus 5.5 to estimate the slopes from the chart, and he calls his own analysis quick.
What the curve does next is open, and Ord says so. The data fit either a continued climb toward human-level scaling or a plateau near an exponent of two thirds, and he allows that future models could scale better than people. His listed caveats are that this is one experiment on one set of tasks, that the token counts run far past a model’s context window and force repeated summarising of its own work, that the slope uncertainty is larger than his error bars show, and that time horizons are an indirect stand-in for human performance.
The second half of the essay concerns recursive self-improvement, meaning an AI system getting better at making its successor better. Researchers who model it use a number, r, for the strength of that loop. When r exceeds 1, the models say progress runs away. In the usual setup, r depends on how well research output grows with more parallel workers and with diminishing returns to more research, calibrated on human researchers.
Ord’s proposal is to fold the token-scaling exponent, which he calls gamma, into that number. Running many AI agents in parallel first converts into more tokens, and tokens then convert into longer horizons at a rate set by gamma. With a gamma below 1, he writes, an intelligence explosion is harder than the standard models assume; above 1, easier. He says current AI parallel scaling looks similar to the human version, so the extra factor falls on the token side.
For readers deciding how much reasoning effort to buy per task, the essay implies little directly. It is about a trend in how returns change across model generations on one kind of task, and it says nothing about prices or about your workload. The one transferable point is that, on Ord’s numbers, the return on extra tokens moved by orders of magnitude in under a year, so a reasoning-budget test run on last year’s model says little about this year’s.
Toby Ord, “AI Scaling vs Human Scaling,” tobyord.com, October 9, 2026.
