Does "think step by step" still work in 2026?

Short answer: sometimes, less than it used to, and not for free. On models with built-in reasoning the gain is marginal at best and was measured as a loss on one of them, while every request costs noticeably more time. On models without built-in reasoning it still helps on average, but it also introduces new errors on questions the model would otherwise get right.

What the numbers actually are

The report does not measure "accuracy went up by X%". It measures how often a model clears a threshold across 25 trials of the same question, and reports the change as a risk difference (RD): the change in the proportion of questions meeting that threshold. Three thresholds are used: every trial correct (100%), at least 23 of 25 (90%), and a simple majority (51%).

Models with built-in reasoning

• o4-mini: a small gain, and only at the loosest threshold (RD = 0.056 at the 51% metric, p = .003).
• Gemini Flash 2.5: a decrease. RD = −0.131 at the 100% metric (p < .001) and RD = −0.071 at 90% (p = .016). Asking it to think step by step made it measurably worse at answering consistently.
• Everything else in that group: no significant change.
• Cost: requests took 20 to 80% longer, 10 to 20 seconds in absolute terms.

Models without built-in reasoning

• Sonnet 3.5: significant gains at the 90% metric (RD = 0.136, p < .001).
• Gemini Flash 2.0: gains at 100% (RD = 0.071, p = .039) and at 90% (RD = 0.111, p < .001).
• But: in three of the five models in this group, chain-of-thought introduced a higher likelihood of errors on easy questions the model would otherwise have answered correctly.
• Cost: requests took 35 to 600% longer, 5 to 15 seconds.

There is also a quieter finding that explains much of the rest: many recent models already perform step-by-step work whether or not you ask. For those, asking adds tokens and time and little else.

Why the advice you will find says otherwise

The technique was introduced in 2022, before any model shipped with reasoning built in, and most of the material explaining it has not been revisited since. To take the page currently ranking first for this technique: as of today it cites work from 2022, and the phrase "reasoning model" does not appear on it at all, nor does o3, o4 or Gemini 2.5. That is not dishonesty, it is a page that was correct when it was written and has not been updated for the models people actually use now.

This is the general shape of prompting advice in 2026: confidently repeated, rarely re-measured.

What to do instead

Check whether your model already reasons. If it does, asking for step-by-step work is mostly a tax. Spend the tokens on context instead.

Use it where the working matters, not where the answer matters. Visible reasoning is genuinely useful when you have to audit an answer. That is a different reason from "it makes the answer better".

Test on your own task before you standardise on it. The effects in this one report differ by model and by threshold. There is no single number to carry around.

Treat variance as a first-class problem. Part of what the report shows is not lower accuracy but wider spread: the same prompt landing differently across trials. If that is your symptom, see what to do when the same prompt gives a different answer every time.

What this page does not claim

This is one report, on specific question sets, at 25 trials per question, on the models available when it was run. It is evidence, not a law. It does not say chain-of-thought is useless, and anyone quoting it to say so, including us, would be overreaching. What it supports is narrower and more useful: the benefit is conditional, the cost is real, and the confident universal advice you will find elsewhere has not been checked lately.

Questions

Should I stop using chain-of-thought?

No. On models without built-in reasoning it still helps on average, and on any model it makes the working visible, which is useful when you need to check the answer. What has changed is that it is not free and not universal.

Why do some models get worse with it?

The report found that on three of five non-reasoning models, asking for step-by-step work introduced errors on easy questions the model would otherwise have answered correctly. Reasoning that is not needed is still reasoning that can go wrong.

Does this apply to my model?

Only testing tells you. The measured effects differ by model within the same report, which is the finding: there is no single number for what chain-of-thought does.

More effects checked against their papers · Failure modes and fixes · Technique index