AI coding assistants: 55.8% faster or 19% slower? Both are right

Two headline studies on AI-assisted development contradict each other:

• Microsoft/GitHub/MIT (2023): 55.8% faster
• METR (2025): 19% slower

Both are accurate. They measure different tasks, different developers, and different definitions of "faster".

Where the speedup is real

The 55.8% came from a single, self-contained, greenfield task with no existing codebase. Google's enterprise RCT (96 engineers, maintenance work in an existing repo) found ~21% faster unadjusted, but the result was not significant after covariate adjustment.

Where it disappears

METR studied 16 open source maintainers across 246 real tasks in codebases they knew well. They finished 19% slower, while believing they were about 20% faster. A longitudinal study at NAV IT found no significant change in commit activity, though developers felt more productive.

At team level

DORA 2024 reported individual gains but worse delivery stability and throughput. DORA 2025 flipped: throughput now correlates positively with AI adoption. AI amplifies what a team already does well or badly.

Practical takeaways

1. Judge work task by task: bounded and greenfield vs. complex and familiar.
2. Measure cycle time and delivered output, not developer sentiment.
3. Hire and grow for reviewing and improving AI-generated code.
4. Fix existing productivity problems (context switching, unclear priorities) independent of the tools.

Full breakdown with the study comparison table: How Much Faster Are AI-Assisted Developers?