No general 10x productivity gain has been established. Studies of AI coding assistance report different outcomes: a large speed gain on one timed programming exercise, a smaller increase in completed workplace tasks, and slower task completion in a trial with experienced developers working in familiar repositories. Those results do not measure the same thing, so none is a universal multiplier for software development.
What did the studies actually measure?
“10x more productive” sounds precise, but productivity can mean several different things: finishing a defined task sooner, completing more tasks during a period, producing code that passes tests, or delivering software that remains correct and maintainable. These measures are not interchangeable. A faster result on one exercise does not by itself establish greater long-term output or quality across a developer’s job.
As an Amazon Associate I earn from qualifying purchases.
The table summarizes the outcomes reported in the studies. Their estimates should be read in the context of each experiment, not averaged into a single AI productivity score.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problems| Study and setting | Participants and work | Reported outcome |
|---|---|---|
| Microsoft Research’s controlled JavaScript exercise (February 2023); GitHub’s account of the same experiment (2022, updated May 21, 2024) | 95 professional developers were randomly assigned Copilot access or no access. They implemented a JavaScript HTTP server; a test suite scored correctness and completeness. | GitHub reported average completion times of 1 hour 11 minutes with Copilot and 2 hours 41 minutes without, and described the Copilot group as 55% faster (95% confidence interval: 21% to 89%; P=.0017). Microsoft Research reports 55.8% faster. These are two accounts of the same experiment, not independent results. |
| GitHub Technical Preview survey (reported in 2022; account updated May 21, 2024) | More than 2,000 developers who had signed up for the preview answered questions about their experience. | Depending on the statement, 60–75% agreed with selected positive statements about fulfillment, frustration, or focus; 73% said Copilot helped them stay in flow, and 87% said it preserved mental effort on repetitive tasks. These are self-reported perceptions, not measured increases in completed work. |
| Three randomized workplace field experiments, summarized by Microsoft Research and published online in Management Science on February 27, 2026 | Pooled data covered 4,867 developers at Microsoft, Accenture, and an anonymous Fortune 100 company. | Access to an AI coding assistant was associated with a 26.08% increase in completed tasks, with a standard error of 10.3%. The individual experiments were noisy and their results varied. |
| METR randomized trial (2025) | 16 experienced open-source developers completed 246 tasks in mature projects where they had an average of five years’ prior experience. The tools were those available from February to June 2025; participants primarily used Cursor Pro and Claude 3.5/3.7 Sonnet when AI was allowed. | Allowing AI increased task completion time by 19%. Participants had forecast a 24% time reduction before the tasks and estimated a 20% reduction afterward. |
| METR follow-up experiment update (February 2026) | The later experiment included 57 developers, 143 repositories, and more than 800 tasks. | METR said selection effects and unreliable time measurements for some participants using multiple agents made the experiment an unreliable signal of the current productivity effect. It did not establish a dependable new estimate. |
Why the 55.8% result is not a 10x finding
The Copilot experiment measured how quickly participants completed one bounded programming task, with correctness and completeness checked by tests. It is evidence that AI assistance helped participants in that particular exercise. It does not establish that they completed every kind of development work at that rate, or that their overall work output multiplied by the same amount.
#1 Best Overall
The result is also not a tenfold gain. “55.8% faster” describes the reported relative difference in completion speed for that task; it does not mean ten times as much finished software. The separate GitHub survey speaks to perceived focus and effort, not a causal change in output. A developer may value that experience, but it cannot be converted into a productivity percentage.
What the workplace experiments add
The pooled field result is closer to ordinary work than a self-contained coding exercise because it came from three workplace experiments. Its measure, however, was completed task count—not task completion time, software quality across all work, or long-run delivery value. The reported estimate is an average across the pooled experiments, which the researchers describe as noisy and varied; it should not be treated as a guaranteed gain for an individual developer or another team.
The study also found that less experienced developers adopted the assistant more and had greater productivity gains. That helps explain why a pooled average may not describe every participant equally, but it does not establish that experience alone predicts whether AI will help in a different organization or task mix.
Recommended Free Tools
Why METR found a slowdown
METR examined a distinct setting: experienced open-source developers modifying mature repositories they already knew well. In that context, allowing the early-2025 tools tested was associated with longer completion times, even though participants expected and later believed they had worked faster.
This is a relevant counterexample to claims that AI necessarily speeds every developer. It is not proof that AI always slows developers: the trial was small and specific, and METR says experimental artifacts cannot be ruled out entirely, while reporting that the result was robust across its analyses. The tools, participants, repository familiarity, and tasks differ from the other studies.
Does METR’s 2026 update settle the question?
No. The later experiment was intended to study a larger, more varied group, but METR said the design could not provide a reliable current effect estimate. Some developers were reluctant to participate if they could not use AI, creating possible selection bias; using multiple concurrent agents also made some participants’ time difficult to measure. The raw estimates therefore should not be treated as a clean resolution of the earlier trial or as a dependable new baseline.
How to judge a productivity claim for your own work
Before comparing percentages—or deciding whether an AI assistant is helping—pin down what “productive” means for the work in question. A useful comparison should make the outcome and conditions explicit:
- Define the work: distinguish a short, self-contained coding task from feature development, debugging, code review, or maintenance in a familiar system.
- Measure the full task: include time spent prompting, checking suggestions, correcting errors, testing, and integrating the change, rather than counting only code generation.
- Keep quality in view: a faster completion time is not a useful gain if correctness, completeness, or maintainability falls.
- Separate output from experience: focus, satisfaction, or reduced mental effort may matter, but they are not the same measure as completed work.
- Match the comparison: note who did the work, their experience and repository familiarity, which tools were available, and the period studied.
What can be concluded about “10x”?
The cited evidence supports a narrower conclusion: AI coding assistance can improve measured outcomes in some settings, but results depend on the task, participants, context, tools, and measurement. A substantial gain on one timed exercise, a pooled increase in workplace task counts, and a slowdown in a familiar-repository trial can coexist because they answer different questions.
None of these studies demonstrates a general tenfold increase in developers’ long-term, quality-adjusted software delivery. That does not rule out a developer achieving a dramatic gain on a particular task; it means the evidence here does not justify presenting 10x as the typical or established effect.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
