The benchmarks the industry leaned on have saturated, which makes progress look like a plateau. On task length — the least saturated measure available — the reach of frontier models has gone on doubling roughly every four to seven months. The pre-training recipe that built the 2023 models has hit real limits; a different engine is now doing most of the work.
The most-quoted way to track AI progress is the score on a hard benchmark, and by that measure the top models have bunched up. On GPQA Diamond, a set of graduate-level science questions where subject-matter PhDs score about 65% and random guessing yields 25%,3 GPT-4 managed 39% in 2023.3 By early 2026 the leading models from OpenAI, Anthropic and Google were all clustered between roughly 91% and 94% — on a 198-item test whose top scores now differ by only a handful of questions.8
Two things are happening at once. The steep early climb is real; the near-vertical jump on the OpenAI line in late 2024 is the arrival of “reasoning” models — o1, which is trained to work through a problem step by step before answering.5 But the flattening at the top is largely an artefact of the ruler, not the runner: once models answer almost every valid question, there is no headroom left to measure, and Epoch AI notes that GPQA and several older tests are now effectively saturated as tools for separating frontier systems.8
To see past the ceiling you need a measure with room to grow. The research group METR uses one: the task-completion time horizon — the length of job, measured by how long a human expert takes, that a model can finish at a given success rate.1 It converts an abstract score into something concrete (“this system reliably does tasks that take people about X hours”) and, crucially, it does not saturate.
Across models from 2019 to 2025, METR found the 50% time horizon doubling about every seven months, an exponential fit with no sign of a plateau.2 Its January 2026 re-estimate put the long-run doubling at 196 days — but the trend measured only on 2023-onward models is faster, around 131 days (~4 months), and the 2024-onward slice faster still.1 A four-month doubling is roughly a sevenfold increase in reach per year; the seven-month rate is about 3.6-fold. Either way the direction is up, and the recent slope is steeper, not shallower.
This is the honest answer to whether the current model set is slowing: on the best unsaturated measure available, it is not — through mid-2026 the pace has held or quickened. But the instrument is straining. METR’s longest calibrated tasks run to about eight hours, so estimates above that are extrapolations, the confidence bands are very wide, and the group flags that measurements beyond roughly 16 hours are unreliable with its current task suite.1 The measure also leans heavily on software and research-engineering tasks, which the labs optimise for hardest.
The plateau intuition is not baseless. The specific thing that drove the 2020–2023 leaps — training ever-larger models on ever-more web text (“pre-training scaling”) — has run into diminishing returns. At NeurIPS in December 2024 Ilya Sutskever, OpenAI’s former chief scientist, argued that “pre-training as we know it will end” because the field has reached “peak data”: compute keeps growing but the single dataset big enough to train on, the public internet, does not.10 OpenAI researchers have said a shortage of fresh data, more than compute, constrained their GPT-4.5-class training run.4 As the analyst Nathan Lambert puts it, both stories can be true at once: scaling still works at a technical level while the rate of visible improvement for everyday users slows.11
The reason the capability curve kept climbing while pre-training stalled is that the industry changed its engine. Chain-of-thought prompting (2022) and then reasoning models such as o1 (September 2024) opened a second axis of scaling: spending more computation at the moment of answering — “test-time” or “inference-time” compute — rather than only at training.9 That shift, paired with reinforcement learning on verifiable tasks, is what the GPQA jump and the lengthening METR horizons largely reflect. In other words, part of the follow-up question — whether a new basis will be needed — has already been answered on the ground: a new basis — in training process, and increasingly in data source — is being adopted now.
On the data side, the constraint is concrete. Epoch AI estimates the usable stock of public human text at about 300 trillion tokens and projects that training sets will grow to meet it somewhere between 2026 and 2032, with a median of 2028.4 The proposed ways round the wall are the other half of the “new basis” question: synthetic (model-generated) data, especially for reasoning; other modalities such as image, audio and video; and squeezing more from the same data.4 Whether any of these fully substitutes for fresh human text is unproven.
Whether the recent, faster doubling rate is a durable regime or a temporary burst from reinforcement-learning gains that will revert toward the seven-month trend. Whether reasoning and test-time compute face their own diminishing returns as pre-training did. And how much of the measured acceleration is real capability versus an artefact of a measuring instrument that runs out of long-enough tasks. The projections here — the METR wedge and the data-wall window — are scenarios on current trends, not forecasts we endorse.
Credible researchers genuinely disagree. This is largely an empirical dispute — the kind evidence can eventually settle — about how long the present paradigm keeps delivering.
Test-time compute and reinforcement learning have opened a second scaling axis with headroom left; the 2024–25 gains came largely without proportionally more training compute, and the METR horizon has, if anything, accelerated. Anthropic’s Dario Amodei has said scaling is “probably going to continue”; OpenAI’s Sam Altman maintains “there is no wall.”11
Pre-training’s data source is finite and near-exhausted; synthetic data has not clearly closed the gap; and inference-time scaling may hit its own diminishing returns, since buying performance with ever-more compute per answer gets expensive fast. Sutskever’s “peak data” and Epoch’s data-wall work point the same way.104
Bottom line. The evidence to mid-2026 supports a narrow claim: the current model set has not slowed on the best available capability measure, and a genuine change of basis — reasoning and test-time compute — is what kept it moving. The broader claim, that this pace continues for years, is not established. The strongest objection to “no slowdown” is that the one metric still showing acceleration is also the one whose measuring instrument is closest to breaking.
METR’s 50% time horizons (Time Horizon 1.1 method), the measured series in Exhibit 02. Values are minutes of human-expert task time; the reader can check the log-scale plot against them.
| Model | Released | 50% horizon (min) | \u2248 human time |
|---|---|---|---|
| GPT-4 | Mar 2023 | 3.5 | ~4 min |
| Claude 3.7 Sonnet | Feb 2025 | 60 | 1 h |
| o3 | Apr 2025 | 121 | ~2 h |
| Claude Opus 4 | May 2025 | 101 | ~1.7 h |
| GPT-5 | Aug 2025 | 214 | ~3.6 h |
| Claude Opus 4.5 | Dec 2025 | 320 | ~5.3 h |
Threads held back to keep this to length: