More capable language models can perform more complex tasks without externalized chain-of-thought reasoning.
Assessment
The evidence, examined directly, establishes the claim.
The claim is comparative: more capable language models can do harder tasks when they are not allowed to reason out loud. Measured evidence supports the direction strongly. The most systematic study, by Ward, Gould, Woodruff and colleagues (June 2026), tested 14 models from GPT-2 to GPT-5.5 on 43 benchmarks with visible reasoning suppressed. It found that the no-chain-of-thought time horizon of frontier models has grown across generations, doubling roughly once a year (373 days, 95% interval 167 to 691). Narrower single-forward-pass math measurements by Redwood Research and METR point the same way. Direct-answer accuracy on standard benchmarks has also risen sharply with scale.
The open questions concern rate and steadiness, not direction. The doubling time is imprecise, the trend rests on a few models, and most of the systematic measurement comes from one research group. Part of the gain may come from recalling memorised knowledge rather than doing internal reasoning, so the claim is better established for "harder tasks" in general than for "longer latent reasoning" in particular. The absolute horizons are still short: a few minutes of human-equivalent work for the best models.
Full reasoning: the evidence and decisions behind this verdict
This pass was triggered because the subclaim on measured no-CoT time horizons received its first assessment: supported, credence 0.85. Its reservations are about the word "steadily", the width of the doubling-time interval (167 to 691 days), and how much of the evidence comes from one group. None of these touch the direction of the effect, and direction is all this parent claim asserts. The comparative claim is also supported apart from the horizon studies. Frontier models answer questions directly at far higher accuracy than GPT-2-era models did, on arithmetic, multi-hop QA and knowledge benchmarks alike. METR's Frontier Risk Report (metr.org/blog/2026-05-19-frontier-risk-report/) and Greenblatt's single-forward-pass math series (blog.redwoodresearch.org/p/measuring-no-cot-math-time-horizon) both agree on direction. The main study is summarised at blog.redwoodresearch.org/p/estimating-no-cot-task-completion. Remaining doubts: the claim's breadth; recall confounding reasoning; reasoning-specialised models that may underperform without CoT relative to their general capability; and elicitation effects. These limit the rate and the interpretation, not the comparative direction. The status stays verified and credence is 0.95. What would change this: evidence that no-CoT gains vanish once memorisation is controlled for, across independent task sets.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because the difficulty of tasks frontier models complete with reasoning tokens suppressed has risen steadily across model generations, and later frontier models are more capable than earlier ones on general benchmarks, more capable models can complete more complex tasks without an externalized chain of thought.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 supported
Granting its premises, the conclusion very probably follows. The step from later generations to more capable models is settled, because frontier capability has risen steadily over the period. The argument rests on the measured growth of no-CoT time horizons. That premise is now assessed as supported for direction, with less certainty about rate and steadiness. This conclusion needs only the direction. The argument stops short of certainty because the horizon metric mixes recall with reasoning and depends on how well reasoning suppression is elicited.
Provenance
Where this claim has been said, linked to its canonical form.
Frontier models can do non-trivial amounts of reasoning without any CoT, and the amount of no-CoT reasoning they can do is increasing exponentially.
Conclusion of a study measuring 14 models (GPT-2 to GPT-5.5) on 43 benchmarks with reasoning suppressed; no-CoT 50% time horizon doubles roughly yearly.
Assessment history
0 status changes over 2 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.