The idea

Use AI coding assistants to get more done as a software developer.

Who the evidence applies to: Working software developers, with the clearest measured gains for less experienced developers and smaller gains or slowdowns for experienced developers on complex, familiar projects.

New studies in 2026, including METR's revised trial and research on agentic coding tools, have revived debate over how much AI coding tools really speed developers up.

How we checked

This is an evidence check: we looked for large studies, official statistics and published datasets, opened every source, and had a second researcher check each figure against them. 5 sources, listed below.

Costs

Start-up cost
Varies
Effort
low

What the evidence shows

The evidence on AI coding assistants points both ways. Field experiments at Microsoft, Accenture and a Fortune 100 company with 4,867 developers found a 26.08% increase in completed tasks, with bigger gains for less experienced developers, and an earlier GitHub Copilot experiment found developers finished a set task 55.8% faster. But METR's 2025 randomised trial found 16 experienced open-source developers took 19% longer with AI tools, despite believing they had been sped up; METR's late-2025 follow-up produced estimates ranging from slower to faster with wide uncertainty. A 2026 study of over 100,000 developers found far larger rises in code written than in releases shipped.

What to expect: Many developers, particularly less experienced ones, are likely to produce more code with these tools, but experienced developers on large, familiar codebases may see little or no speed-up, and more code does not translate one-for-one into more finished work. The studies measure output, not pay.

First steps

  1. Try an AI coding assistant on a contained task and time yourself with and without it rather than relying on how fast it feels
  2. Review, test and understand any generated code before merging it
  3. Use it most where you are less familiar, such as a new language or boilerplate, and judge separately whether it helps on code you know well
  4. Track finished outcomes such as merged pull requests or releases, not just lines of code

Limitations

Studies report averages: your own result depends on your field, skills and where you live. Figures come from the sources below and may be updated as new evidence appears.

VerdictWorks with limits

Works for some, not most.

Sources

  1. Cui, Demirer et al., randomised field experiments at Microsoft, Accenture and a Fortune 100 company with 4,867 developers: 26.08% increase in completed tasks; less experienced developers adopted more and gained more economics.mit.edu
  2. Peng et al., GitHub Copilot controlled experiment: developers given Copilot implemented an HTTP server in JavaScript 55.8% faster than the control group arxiv.org
  3. METR randomised trial (July 2025): 16 experienced open-source developers took 19% longer with AI tools, having expected a 24% speed-up and still believing afterwards they were 20% faster metr.org
  4. METR update (February 2026): late-2025 estimates of an 18% speed-up for returning developers and 4% for new ones, both with confidence intervals including a slowdown; METR says selection effects make results unreliable metr.org
  5. Demirer, Musolff and Yang on VoxEU (June 2026), over 100,000 developers: AI tools greatly raise code output, but a more than sevenfold rise in lines of code from sync agents became a 65% rise in pull requests and only 20% more releases cepr.org