A research essay published by Forethought on 9 November splits “test-time scaling” into two different things, and argues the less discussed one will matter far more than the first. The essay is an argument, not a measurement, and its author says plainly that the key evidence is circumstantial.
Test-time scaling means giving a model more computing power and more time on a single question once it is already deployed, usually by letting it write a longer chain of reasoning before it answers. The familiar version has a ceiling. A model trained with reasoning chains of up to 100,000 tokens gets better as you let it think toward that length, and then stops getting better. Forethought’s essay says the earliest reasoning models, o1 and R1, behaved this way.
The second version has no such wall. Here the model keeps improving when you hand it millions or billions of tokens, far beyond anything it practiced on. The essay calls this “compute extrapolation” and measures it with two numbers: the ceiling, meaning how many multiples of the training budget help before gains stop, and the slope, meaning how fast those gains taper off.
The clearest illustration is old. AlphaGo Zero, the Go program, got about 0.4 seconds of thinking per move in training, 5 seconds in one evaluation, and roughly a minute in a tournament. That is a ceiling of at least 100 times. The essay estimates that without this trick, training the program to the same level would have cost 10 to 50 times more, and it believes language models are now getting the same discount.
Why it matters comes down to who can afford to be ambitious. If a lab wants to spend an enormous sum of computing on one consequential task, such as finding better machine-learning algorithms, it no longer has to have spent that sum on many practice tasks first. A single expensive attempt at test time is cheap next to a training run. The essay expects the spending race to move from training to inference, with millions of dollars poured into single tasks and plenty of failed bets.
The second consequence lands on safety testing. To check whether a model can do something dangerous, evaluators have to give it as much computing as a determined user might. Without extrapolation, that budget is capped by the training budget. With it, the budget has no obvious limit, and the essay says evaluations may only be able to show that a danger is absent below some spending level, and only in a handful of high-priority areas. It cites the UK AI Security Institute’s earlier finding that good evaluations of smarter models already cost more inference compute.
The evidence is the weak spot, and the essay says so. Anthropic and OpenAI do not publish how much computing their training tasks use, so outsiders cannot observe extrapolation directly. The strongest data comes from an academic paper, “The Markovian Thinker,” in which a model trained to think in chunks, discarding old reasoning as it goes, kept improving at 30 or more chunks after training on at most five. A model trained the standard way, with 24,000 tokens of unbroken thought, did not.
For the commercial models, the author points to indirect signals around the end of 2025 and early 2026. Toby Ord’s analysis of UK AI Security Institute data suggests Opus 4.5 needed 370 times more reasoning tokens to match what a human gains from thinking 10 times longer, while Opus 4.6 needed 46 times. A benchmark called MirrorCode had Opus 4.6 still improving after tens of context compactions on a billion-token budget. The picture for GPT-5.1-Codex-Max, from OpenAI, is murkier: METR reports that compaction did not lengthen the tasks it can complete.
The essay also floats a policy idea: asking labs to disclose their most compute-hungry training tasks, which would give outsiders a baseline. Until that exists, any product team pricing an agent feature on the assumption that more thinking time buys proportionally more quality should test that curve directly on its own hardest tasks, because the vendor’s training budget no longer tells you where the gains end.
Reported by Forethought in its research essay “Two Kinds of Test-Time Scaling,” published 9 November 2026.
