You will hear claims from "55% faster" to "it slows us down". Both can be true, because they measure different things.
Controlled task experiments
In a 2023 experiment, developers given GitHub Copilot implemented an HTTP server in JavaScript 55.8% faster than a control group. It is a clean experiment — random assignment, one well-defined task — but the task was greenfield and self-contained, unlike most work in a large codebase.
Field studies of real work
In 2025, METR ran a randomized study with 16 experienced open-source developers working on 246 real issues in large repositories they knew well. With early-2025 AI tools (mainly Cursor with Claude 3.5/3.7 Sonnet), they took 19% longer. Before starting they expected a 24% speed-up, and afterwards they still believed AI had sped them up by 20%. METR later noted the results are out of date as tools improved.
Large telemetry and survey datasets
Engineering-intelligence vendors such as DX combine tool telemetry with surveys across hundreds of companies. DX's Q4 2025 report (135,000+ developers, 435 companies) found developers report saving 3.6 hours a week on average. By its Q1 2026 report, 93% of developers were using AI tools: adoption, which on its own says nothing about delivery.
How to read any claim
- Task or system? Speed on a task is not the same as faster delivery of features.
- Self-reported or measured? METR shows perception and reality can point in opposite directions.
- Whose code? Greenfield, legacy, or a mature codebase experts know well.
- Which tools and when? Results age quickly.
- What about quality? A speed-up that raises change failures is not a gain.
The most reliable evidence for your organization is your own: a baseline, a pilot with a comparison group, and the same metrics afterwards.