They felt 3x faster with AI. The clock said 19% longer.

The median technical worker in METR’s 2026 survey says AI has tripled their speed. The figure is self-reported, and when a pitch deck reaches for it, it arrives sounding like a measurement. It is not one. The survey that produced it says so, the one experiment that tested it ran the other way, and the follow-up to that experiment is now, by its own authors’ account, unreliable. For anyone raising or writing a first cheque this year, the story of this one number is a short course in what early-stage claims are made of.

A median with a warning attached

The survey is careful about what it is. METR reached 349 technical workers: 87 software engineers, 71 researchers, 129 academics and PhD students, 48 founders and managers. The response rate was roughly 2% of those emailed, and about 70% of participants were paid to take part. This is a self-selected sample of people interested enough in AI tooling to answer a long survey about it, and METR says as much.

Two findings sit side by side. The median self-reported speed change was 3x. The median self-reported change in the value in their work was 1.4-2x. The speed figure is the one that runs highest, and METR expects exactly that, in the same sentence that reports it:

The median self-reported speed change (which we expect to be higher than value change) is 3x.

One paragraph later, METR says it outright:

Importantly, survey results are not necessarily grounded in reality.

The one measured test

METR has checked feelings against a clock exactly once. In early 2025 it ran a randomised study with 16 experienced open-source developers working 246 real issues on repositories they knew well. The result became famous, and deserved to:

When developers are allowed to use AI tools, they take 19% longer to complete issues

The developers had forecast a 24% speed-up before starting. After finishing slower, they still put their gain at 20%. METR’s own summary, which its 2026 survey post still cites, is that people overestimated AI’s effect on their time by 40 percentage points on average.

That is the calibration lesson: the gap between what the work felt like and what the clock said was not a rounding error. It was the whole claim, with the sign flipped.

The February correction

Whoever cites the 19% today is citing a number its own authors have marked out of date. The study page now opens with a banner:

These results are out of date. We have released results that are current as of early 2026, in a continuation of this study.

In February 2026 METR changed its experiment design and reported that the new data gives an unreliable signal of the current effect of AI tools. For the developers who returned:

For the subset of the original developers who participated in the later study, we now estimate a speedup of -18% with a confidence interval between -38% and +9%.

An interval that wide is the shape of current knowledge. METR believes developers are likely faster with the tools now than in early 2025, and says why the size stays unknown:

However, because of the selection effects in our experiment, our data is only very weak evidence for the size of this increase.

Selection effects, in plain terms: those who came back are not a random draw of those who started. Read the three documents together and the position is stark. The felt number runs high, the measured one is historical, and the organisation that produced both now calls its current data unreliable. METR is left with no dependable multiplier to hand an investor, and it says so.

What this means at a first cheque

A first cheque is underwritten on claims, because operating history is thin and the accounts, where they exist, are young. What separates one claim from another at that stage is not confidence. It is method.

So the diligence question worth asking about any AI efficiency line costs nothing: measured how, against what baseline? Most conversations end there. A founder can be the exception. Time one core workflow with the tools and without, run it more than once, and put the ratio in the deck with the method and the spread beside it. The number may come out modest. Modest with a method beats triumphant without one.

For an angel, the same question is portfolio hygiene, not scepticism. A deck line built on a felt multiple is not a lie. In the one case METR put on a clock, the people doing the work overestimated their own speed by 40 points without noticing. The question moves the conversation onto ground where evidence can exist, and it does so without costing the relationship anything.

What changes now is small and specific. An investor who asks how a speed-up was measured is no longer being difficult; METR’s own record is the licence to ask. A founder who answers with a stopwatch number, however modest, is showing something the published evidence cannot currently show. Angels Den has introduced founders to investors at the first cheque since 2007. The cheques are still written on claims. The question decides which claims are worth the ink.

Sources. METR, Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity, metr.org, May 2026. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025, since marked out of date. METR, We are Changing our Developer Productivity Experiment Design, February 2026. All figures re-verified against the live pages in July 2026; the retrieval log sits in the cycle fact-lock.