AI Coding Productivity Statistics 2026

A 16-developer METR trial measured 19% more task time with AI allowed, while trials run by AI tool vendors and developer surveys reported gains. We compare effect sizes by study design, self-reported and measured productivity, and results by affiliation.

Published October 24, 2026

In July 2025, the research nonprofit METR found that experienced open-source developers took 19% longer to finish tasks when AI tools were allowed in a 16-developer randomized trial, while trials run by AI tool vendors and developer surveys report gains. That split led us to study how much AI coding tools change developer output in published trials, field experiments, telemetry studies and surveys, and how strong each piece of evidence is.

Our analysts compiled 10 studies published from February 2023 to February 2026. The set includes three randomized controlled trials, one set of randomized field experiments at three companies, three observational telemetry studies and three developer surveys. We coded each study by design, sample size, population, metric, measured effect and the affiliation of its authors or publisher, using the rules in the methodology section. Where one study reported both what developers believed and what researchers measured, we set the two figures side by side.

In the following sections, we code every study, compare effect sizes by study design, look at the task-time trials and the company field experiments, review telemetry results on output and code quality, cover the DORA estimates and the 2024 and 2025 developer surveys, compare self-reported and measured productivity, and group the evidence by affiliation.

How We Coded Each AI Coding Productivity Study

This section lists every study in our dataset with the codes we assigned to it. Design, sample, population, metric and effect come from each source, and the affiliation code follows the employer of the authors or the organization that published the work.

Source: AI Services Companies, AI Coding Productivity Study, October 2026. Uplevel's 785 is our calculation (351 with Copilot access plus 434 without). Data: ai-coding-study-coding.csv

• The 10 studies measure task time, output counts, code quality, delivery or perception, and a figure from one class cannot stand in for a figure from another. Each effect needs its metric beside it when it is cited.

Six of the 10 studies had authors employed by an AI tool vendor or were published by a vendor (our count, n=10 studies). Microsoft and GitHub employees wrote or co-wrote two of them, Google ran or wrote three, and Stack Overflow, which offers an AI-powered VS Code extension with its Stack Internal plans, published one. The four randomized studies together covered 5,074 developers (our calculation, 16 + 95 + 96 + 4,867), and 95.9% of them came from the field experiments (our calculation, 4,867 of 5,074). Among the three randomized controlled trials, METR's is the only one in which developers worked on issues from projects they already maintained. The Peng and Google trials gave every participant the same assigned task.

AI Coding Productivity Effect Size by Study Design

This section groups the 10 studies by design and compares the headline effects each group reported. Task-time trials report a percent change in time to finish a task, field experiments and telemetry report changes in output or quality, and surveys report what share of respondents perceived a gain.

Source: AI Services Companies, AI Coding Productivity Study, October 2026. Bases, counts and directions are our calculations. The survey base adds DORA's approximate counts of 3,000 and 5,000 to Stack Overflow's 33,662. Data: ai-coding-effect-by-design.csv

• The designs that reach the most people rest on the weakest evidence of cause. Claims about speed have to come from the small randomized trials, and claims about scale come from surveys and telemetry that cannot separate the tool from everything else that changed.

Measured speed effects in the three task-time trials span 74.8 percentage points between the largest cut in task time and the only increase (our calculation, n=3 trials). Two of the three telemetry studies measured throughput and split on it, and all three reported worse results on at least one quality measure. Every outcome in the surveys is the respondent's own report, including the DORA delivery measures.

Randomized Trials of AI Coding Tools on Task Time

This section compares the three trials that timed developers on tasks with and without AI assistance. Each trial assigned AI access at random, and their tasks and participants differ widely.

METR recruited developers who had contributed to large open-source repositories for an average of five years, with the repositories averaging 23,000 stars. Each developer supplied a list of real issues from their own project, and each issue was randomly assigned to allow or disallow AI before work began, giving 136 completed issues with AI allowed and 110 without. Developers mainly used Cursor Pro with Claude 3.5 and 3.7 Sonnet between February and June 2025, and METR paid them $150 per hour. While 93% had used large language models before the study, 44% had prior experience with the Cursor editor (n=16 developers).

Source: METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (July 2025); Peng, Kalliamvakou, Cihon and Demirer, The Impact of AI on Developer Productivity (February 2023); Paradis et al., How Much Does AI Impact Development Speed? (2024). Data: ai-coding-task-time-rcts.csv

• Each trial's speed figure describes its own setting. A result from one assigned task in an unfamiliar codebase says little about maintainers working through their own backlog, and the reverse holds too.

The GitHub Copilot experiment by Peng and colleagues sent 166 offers through Upwork and randomized the 95 developers who accepted. The authors report that thirty-five developers from both groups completed the task and survey, and the timing comparison uses the developers who completed it. Treated developers finished the JavaScript server task in an average of 71.17 minutes, against 160.89 minutes for the control group, so the control group took more than twice as long (our calculation). Most participants were aged 25 to 34 and lived in India or Pakistan, and the median income band was $10,000 to $19,000. The treated group's success rate was 7 percentage points higher, a difference the authors report as not statistically significant.

Google's trial enrolled 96 full-time engineers in summer 2024 and tested three AI features in its internal development tools. Engineers had to have worked at Google for at least a year and be proficient in C++, and the task asked them to add a data logging component in code they were not familiar with. The mean time on task was 104.4 minutes, and 93 of the 96 participants who started the task completed it (96.9%). The authors estimate that AI cut time on task by about 21%, with a p-value of 0.086, which does not meet the 5% threshold for statistical significance. Of the three trials, Peng's is the only one with a speed gain whose confidence interval excludes zero.

Field Experiments at Microsoft, Accenture and a Fortune 100 Company

This section covers the largest randomized evidence in our dataset, three field experiments in which companies randomized which developers got GitHub Copilot access, or when they got it, during their normal work. Cui, Demirer, Jaffe, Musolff, Peng and Salz report the results in a paper Management Science published online in February 2026.

The Microsoft experiment, involving 1,746 developers primarily located in the United States, began in September 2022 and ended earlier than planned, in April 2023, when control-group developers began to request access. Accenture's experiment started in the last week of July 2023 in a number of offices in Southeast Asia, assigned 61.3% of its 320 developers to treatment and included a training session. Its control group went five months without access, until December 2023. At the anonymous Fortune 100 company, every developer eventually got access, and the experiment randomized when developers were invited to GitHub Copilot, with weekly invites in September and October 2023.

Source: Cui, Demirer, Jaffe, Musolff, Peng and Salz, The Effects of Generative AI on High-Skilled Work, Management Science (online February 2026). Figures from the 2025 working paper version. Effects are weighted IV estimates expressed as a percentage of the control mean. The assigned total is our calculation (1,746 + 320 + 3,054). Data: cui-field-experiment-results.csv

• These experiments count work submitted, in pull requests, commits and builds. Their only significant quality result was a decline, so the output gain comes with limited evidence on whether the extra work held its quality.

The analyzed samples are smaller than the assigned groups because the authors kept only software developers at Microsoft and dropped those who switched organizations or adopted Copilot before the experiment, dropped Accenture developers who left the company, and dropped 24 early adopters at the anonymous company. Because compliance with assignment was imperfect, the authors use assignment to treatment as an instrument for Copilot adoption and weight the estimates across experiments. Pooled commits rose 13.55%, which was not statistically significant, and pooled builds rose 38.38%, significant at 1% (n=4,867 developers).

Two of the six authors work at Microsoft, which owns GitHub. The experiments were registered with the AEA RCT Registry as AEARCTR-0014530 after they began.

Telemetry Studies of AI Coding Output and Code Quality

This section reviews three studies that measured AI coding effects from engineering system data without random assignment. Their bases are larger than the trials', and each was published by a company that sells software for measuring engineering work.

Uplevel compared 351 developers with Copilot access and 434 without across several enterprise customers, using January 9 to April 9, 2023 as the before period and January 8 to April 7, 2024 as the after period. It grouped developers by whether they had access, and it describes its results as observational.

Source: Uplevel Data Labs, Can Generative AI Improve Developer Productivity? (September 2024); Faros AI, The AI Productivity Paradox Report 2025 (July 2025); GitClear, AI Copilot Code Quality (February 2025). Uplevel's 785 is our calculation (351 + 434). Data: ai-coding-telemetry-results.csv

• The costs in these studies land after the code is written, on reviewers, testers and maintainers. A productivity measure that counts only what authors produce leaves those costs out of the result.

Uplevel says some of its metrics were statistically significant but that the actual change was inconsequential to engineering outcomes, and it gives the 1.7-minute cut in cycle time as its example. Sustained work outside standard hours fell for both groups over the year, by 17% for developers with Copilot access and by almost 28% for those without.

Faros drew on task management, IDE, static analysis, CI/CD, version control, incident and HR system data from more than 10,000 developers. For each team, it compared the two quarters with the lowest AI adoption against the two with the highest, and it reported only metrics with data from at least six companies and Spearman correlations significant at p<0.05. Team-level gains in Faros's data did not carry through to company-level results. Faros reports that widespread use, above 60% weekly active users, began at most companies only in the two or three quarters before its analysis, that use was highest among engineers newer to their company, and that most developers used only autocomplete features. Developers on high-adoption teams touched 9% more tasks and 47% more pull requests per day.

GitClear classified 211 million changed lines authored from January 2020 to December 2024 into operations such as added, moved, copy/pasted and churned. Its data cover all changed lines in the repositories it tracks, whether a developer or an AI assistant wrote them. GitClear counts a line as churned when it is reverted or substantially revised within two weeks of being pushed, and its 2024 data are the first year in its five-year series in which copy/pasted lines outnumbered moved lines.

DORA Estimates of AI Adoption and Software Delivery

This section reports DORA's 2024 model estimates of how outcomes change as AI adoption rises, and the direction of the 2025 results. DORA measures both adoption and outcomes through its survey, then estimates the change in each outcome for a 25% increase in an individual's AI adoption, using 89% uncertainty intervals.

Source: DORA, Accelerate State of DevOps Report 2024 (Google Cloud). Data: dora-2024-ai-adoption-estimates.csv

• In DORA's model, most individual and process estimates improve with adoption while both delivery estimates worsen. A team can report better flow, documentation and review speed while the same model links its rising adoption to weaker delivery.

Stability carried the largest delivery estimate in the 2024 model, roughly five times the size of the throughput estimate (our calculation). Documentation quality had the largest positive estimate of the 13 outcomes in the table.

DORA's 2025 report, State of AI-assisted Software Development, drew on survey responses from nearly 5,000 technology professionals and more than 100 hours of qualitative data. Google Cloud's announcement of the report says DORA now observes a positive relationship between AI adoption and both software delivery throughput and product performance, and a continued negative relationship with delivery stability. DORA's explanation is that faster change volume exposes weak testing, version control practices and feedback loops downstream, and that teams with loosely coupled architectures and fast feedback loops see gains while teams in tightly coupled systems see little or no benefit.

Developer Survey Results on AI Coding Productivity, 2024 and 2025

This section collects the adoption, perceived productivity and trust figures from the DORA and Stack Overflow surveys. Each figure is a share of respondents to one question, with the base for that question where the publisher gives it.

Source: DORA, Accelerate State of DevOps Report 2024; Google Cloud, Announcing the 2025 DORA Report (September 23, 2025); Stack Overflow, 2025 Developer Survey, AI section. Data: ai-coding-survey-results-2025.csv

• Wide use and wide distrust appear in the same surveys. An adoption figure shows how many developers use the tools, and it leaves open whether their output improved.

Stack Overflow's 52% figure on productivity comes from a question asking whether AI tools or agents changed how respondents complete development work, answered by 31,636 people. In that question, 16.3% said yes to a great extent and 35.3% said yes somewhat. Among all Stack Overflow respondents, the share using or planning to use AI tools rose from 76% in 2024 to 84% in 2025, and favorable sentiment toward AI tools fell from above 70% in 2023 and 2024 to 60% in 2025.

The share of DORA respondents with little or no trust in AI-generated code fell by about 9 points from 2024 to 2025 (our calculation, nearly 3,000 and nearly 5,000 respondents). Among Stack Overflow respondents, distrust of AI accuracy outweighed trust by 13 percentage points (our calculation, n=33,244). The most common frustration in the Stack Overflow survey was output that is close to correct, and debugging AI-generated code came second (n=31,476).

Self-Reported vs Measured AI Coding Productivity

This section compares what developers believed about their speed with what the same study measured. Two studies in our dataset asked participants for a speed estimate and also timed them.

Source: METR (July 2025); Peng, Kalliamvakou, Cihon and Demirer (February 2023). Gaps are our calculations as the difference between the self-reported and measured percentages, with METR's extra time counted as negative time saved. Peng's self-estimate is a productivity gain and its measured figure is a cut in task time, which are different units. Data: ai-coding-self-reported-vs-measured.csv

• METR's developers forecast that AI would cut their task time by 24% and estimated a 20% cut after finishing, while their measured task time rose 19% (n=16). Peng's developers estimated a 35% productivity gain, and the measured cut in task time was 55.8%. A pilot that relies on developers' own sense of speed can overstate or understate the effect.

METR also asked outside experts to forecast the result before the trial. Economics experts predicted AI would cut completion time by 39% and machine learning experts predicted 38%, so both groups expected larger time savings than the developers did. METR's developers still believed AI had sped them up after finishing the tasks, and their post-task estimate differed from their forecast by 4 points (our calculation, n=16).

In the Peng experiment, both the treated and control groups gave an average estimate of a 35% productivity gain, with control participants estimating the gain they would have had with Copilot. The authors describe that estimate as an underestimate of the 55.8% gain they measured, the opposite error from METR's developers.

AI Coding Productivity Evidence by Affiliation

This section groups the 10 studies by the affiliation code from Table 1 and records which designs each group used and what direction its measured results took.

Source: AI Services Companies, AI Coding Productivity Study, October 2026. Affiliation codes and counts are our calculations. Data: ai-coding-affiliation-by-design.csv

• Every randomized study except METR's has at least one author employed by a company that sells AI coding tools. Without an independent trial of the same kinds of tasks, those gains rest on vendor-affiliated research alone and carry less weight than a replicated result.

Three randomized studies found a gain, the Peng and Google trials and the Cui field experiment, and Google's gain did not reach statistical significance. Peng's co-authors came from Microsoft Research, GitHub and MIT Sloan, and all authors of the Google trial work at Google. Two of the six Cui authors work at Microsoft, and the other four are affiliated with Princeton, MIT and the Wharton School. DORA is run by Google Cloud, and Stack Overflow offers an AI-powered VS Code extension with its Stack Internal plans. METR describes itself as a research nonprofit, and Uplevel, Faros and GitClear sell software that measures engineering work. The three analytics vendors used observational designs exclusively.

What AI Coding Productivity Evidence Means Going Forward

AI coding tools can shift effort from the developer who writes the code to the people who review and maintain it.

The tools also change between releases, so each result describes the assistants and workflows its participants used at the time. The evidence so far shows that AI coding tools change how developers write software, and that the size and direction of the effect on productivity differ with the task, the team and the measure.