Remember "AI makes developers 19% slower"? Here’s the 2026 update.

"AI makes experienced developers 19% slower."

The numbers

1. METR’s 2025 randomized trial19%

   longer to finish tasks with AI, for experienced open-source developers.

   Confidence interval: 2% to 39% slower.

   Source: METR, 24 Feb 2026 update (2025 result)

2. Late-2025 follow-up800+

   tasks, 57 developers, 143 repos. The result: no clear answer.

   Estimated speedups of 18% and 4%, but both intervals cross zero.

   Source: METR, 24 Feb 2026

3. Why it broke30-50%

   of developers held back tasks they didn’t want to do without AI.

   METR says this likely makes its speedup estimate a lower bound.

   Source: METR, 24 Feb 2026

4. METR, on its own new data

   “we believe that the data from our new experiment gives us an unreliable signal”METR, Feb 2026

   Source: METR, 24 Feb 2026

5. METR survey, May 2026, 349 technical workers

   Felt speed vs felt value

   Median speed change reported3xself-reported

   Median value change reported1.4-2xself-reported

   In the 2025 trial, developers’ own estimates were off by about 40 points.

   Source: METR AI usage survey, 11 May 2026

Stop quoting one study

1. "19% slower" was early 2025. Tools changed. METR now says that estimate is likely a bad proxy for real productivity.
2. "3x faster" isn’t proven either. That’s the median self-report in METR’s May survey. Self-reports run hot.
3. Measure your own team. Cycle time, review wait, rework. Before and after, same team.

Sources

1. METR, We are changing our developer productivity experiment design, 24 Feb 2026
2. METR, AI usage survey, 11 May 2026

Faster, slower, or just different? Tell me on LinkedIn or Instagram.