AI Coding Tools Speed Up Simple Tasks by 55% and Barely Move Real-World Productivity
Early vendor-backed studies from GitHub, Google, and Microsoft found developers completing tasks 20% to 55% faster with AI assistance. Independent, controlled research tells a messier story: METR's July 2025 randomized trial found experienced open-source developers were actually 19% slower with AI tools even though they believed they were 20% faster, and by early 2026 METR's own updated estimate for the same cohort landed at roughly an 18% speedup with a wide uncertainty band. McKinsey separately found a 46% time reduction on routine coding tasks but under 10% on debugging and system design. The gap between the marketing number and the measured number is the story.
Who has this problem
Engineering leadership (VPs of Engineering, CTOs) at mid-to-large orgs
Leaders who've adopted AI coding tools and need real, independent productivity data before making further tooling or headcount decisions.
Individual developers questioning their own perceived speedup
Engineers who believe AI tools make them faster (per METR, often incorrectly) and want an honest read on their actual output.
Startup founders deciding whether to buy AI seats or hire
People weighing vendor claims of 20-55% speedups against independent results.
Finance and procurement teams approving AI-tool budgets
Approvers who need measured, not marketed, productivity numbers.
Signal timeline
Faros AI publishes its 'AI Engineering Report 2026: The Acceleration Whiplash,' based on two years of telemetry from 22,000 developers across 4,000+ teams, finding real throughput gains under AI adoption alongside a proportionally larger rise in bugs and review burden.
Keep reading with access
You've seen the scores, who has this problem, and the latest signal. Members get everything behind it.
- 6 dated, sourced signals
- 4 pieces of hard evidence
- 2 builders and where each falls short
- 2 charts and tables
- 7 linked sources
Already a member? Log in
Signals are specific dated events, each linked to its source and reviewed before entry. The read is editorial interpretation, not financial or investment advice. Do your own diligence before building or investing.