Faster is what it felt like.

METR randomly assigned 246 real issues across 16 developers who already knew large open-source repositories. These were not lab exercises. They were the kind of work those people would have done anyway, in code they knew. For each issue, the tool was either allowed or it was not. That is the comparison: the same kind of task, with the tool and without it.

When AI use was allowed, mostly Cursor Pro with Claude 3.5 or Claude 3.7 Sonnet, between February and June 2025, the tasks took 19 percent longer. They were slower. The interval on that result runs from 2 percent to 39 percent. It is a wide interval. The 19 percent is the center, not a figure exact to the last point. What the interval does not include is a speedup. Even the low end is still slower.

Before the study, the participants expected to be 24 percent faster. Afterward, with the work finished and the slowdown already behind them, they still believed they had been 20 percent faster. The experience did not correct the belief. People who know their own work reported an improvement the clock did not find. There is no need to assume they were lying. The feeling of going faster and the time measured are not the same fact. A tool that answers quickly on screen can feel fast while the whole task grows longer at a point nobody is watching.

The authors do not generalize. They do not say the result holds for most developers, or for other kinds of work, or for today’s tools. On 24 February 2026 they called the result historical and said it no longer describes the current effect. It is the record of a measurement made between February and June 2025, with those people and those tools. Nothing more, and nothing less.

What remains, and it does not depend on whether the 19 percent still applies, is the distance between what was believed and what was measured. A tool that feels fast still needs the person who checks the result. The time of the whole task has to be measured, not only the feeling it left. The feeling was sincere. It was not the measurement.

For that reason, do not judge a tool by how fast it feels. In that study, the tasks took longer, and the people doing them believed they had gone faster. Measure the whole task, from start to finish, before you trust the feeling. And do not use the 19 percent as if it described today’s tools: in February 2026 the authors said the result was already historical.