AI coding agent usage statistics 2026 — why 90% and 59% are both right

1. Executive summary

First, the terms. An AI coding agent is a tool that goes beyond autocomplete or chat: it reads files, runs commands and carries out multiple steps on its own. Claude Code, GitHub Copilot CLI, Codex and Cursor's agent mode fall into this category. Each survey draws this line differently, and that is the first reason the numbers differ. A pulse survey is a small survey run at short intervals, separate from the annual survey. Merged PRs are the number of pull requests (code change requests) that passed review and were actually merged.

The figures in this article are exactly as stated in each survey announcement and in the original paper. A difference in percentage points (pp) is the difference between two rates. We did not run our own measurements.

2. Before reading the numbers: four survey sources, six datasets, their samples and questions

A "usage rate" means different things depending on which of the sources below it comes from. Every number in this article carries a citation number from this table.

The table makes three things clear right away. First, all of the surveys rely on voluntary responses. Weighting corrects for differences in region and experience, but not for the tendency of people interested in AI tools to respond more often. Second, three of the four survey sources have a stake in the results: JetBrains and Anthropic sell AI coding tools, and Microsoft owns GitHub. Only Stack Overflow does not sell such tools. Third, the timing differs. Stack Overflow's annual survey values are from mid-2025, and JetBrains' are from mid-2026. In this market the tool rankings change within six months, so the two should not be read as a difference measured at the same point in time.

3. Usage rates: what 90%, 68%, 59% and 51% each mean

Read from top to bottom, the table explains the gap between 90% and 14.9%. The 90% asked professional developers in mid-2026 about use "at least once a week"; the 14.9% asked in mid-2025 about "daily" use. Between the two lie a one-year gap, a difference in frequency ("at least once a week" vs. "daily") and a difference in how broadly "agent" was defined. JetBrains says it includes both local agents and cloud agents, while Stack Overflow 2025 asked about agents narrowly, in a separate question. [1] [6]

When the same source reports values at two points in time, the direction is clear. Stack Overflow's comparison has agent use rising from 31% to 59% in a year, and daily use from 14% to 37%. But the 2025 values come from the annual survey of 49,009 people and the 2026 values from the pulse survey of 1,100 people, so the sampling frames differ. Stack Overflow itself put the two values side by side, but we read the comparison only for its direction, not its size. [4]

4. Share of users by tool: the conditions behind Claude Code's 39%

Korean news coverage most often cited the claim that "Claude Code overtook Copilot for first place" (our translation). The figure comes from a single source, JetBrains, and because JetBrains asked the same question three times, it can be read as a time series. [3] [1]

Share of professional developers, based on "regular usage at work." Values through January 2026 are from the AI Pulse survey (more than 10,000 respondents) and the May–July values are from the Developer Ecosystem Survey (more than 15,000 respondents), so the samples differ. A dash means the announcements give no value. For the US, Claude Code was at 24% in January 2026 (US and Canada) and 47% in May–July.

Three things can be confirmed from this table. Claude Code rose from 3% to 39% in a little over a year. Over that period, Copilot fell from 29% to 21% and Cursor from 18% to 12%, while Codex rose from 3% to 16%. JetBrains says 31% of Claude Code users named Claude Code as their main tool, and its January 2026 survey reported a customer satisfaction score (CSAT) of 91% and a Net Promoter Score (NPS) of 54. [3] [1]

We also note what cannot be confirmed. These percentages do not add up to 100%. One person can use several tools, and JetBrains did not publish the share of respondents who use more than one. That is why "share of users" is a more accurate name than "market share." In addition, JetBrains' own product (JetBrains AI, 9%) is on the list, and the weighting reflects familiarity with JetBrains products, so these figures were measured by a party with a stake in the tool rankings. At the time of writing, no other survey source had measured the same question at the same time. Once Stack Overflow's 2026 annual survey is published, a cross-check will be possible.

5. How much gets delegated: use and trust do not move together

A 90% usage rate does not mean agents write 90% of the code. The same JetBrains survey asked separately how code gets written. On average across respondents, about 47% of code was generated by agents, about 38% was written with AI assistance and about 27% was written by hand. The three values add up to more than 100% because we copied the announcement's averages as published; the question structure was not published, so we did not adjust them. More than half of respondents said they write less than 20% of their code by hand, and one in five said they write no code without AI. JetBrains divided respondents into three groups: agentic coders, about 31% (on average 84% of their code agent-generated); AI-assisted coders, about 47% (40% on average); and manual coders, about 23% (about 10%). [2]

The three surveys point the same way. More people use agents, but people still control how much they hand over. In the words of Stack Overflow's pulse survey, "Full autonomy is a risk agentic users are not willing to take." Anthropic's report is a vendor document with no published methodology, so its evidence carries different weight, but it points the same way: AI is used in 60% of work, yet fully delegated work is 20% or less, and what gets delegated first is work that is easy to verify and work that is low-risk. [4] [7]

Trust indicators moved in the opposite direction from usage. In Stack Overflow's 2025 annual survey, usage rose, yet more respondents distrusted the accuracy of AI output (46%) than trusted it (33%), and only 3% highly trusted it. In the pulse survey, the share who strongly agreed that accuracy is a barrier fell from 57% to 47%, but security concerns (44%) remained. Usage rising while trust stays low means the review burden still falls on people. We covered how to review AI-generated code in our report on the quality of AI-written code. [6] [4]

6. Does it lead to more output? The scope of the 24% merged-PR lift

Surveys measure "I use it" and "I feel more productive." In Stack Overflow 2025, about 70% of agent users said agents had reduced the time spent on specific tasks and 69% said agents had increased their productivity, but these are self-reports. Only one study measured actual output from usage records: Microsoft's. It is a preprint published on July 1, 2026, and all three authors work at Microsoft. [5]

The study has two parts. The adoption study used the records of tens of thousands of engineers to look at who tries Copilot CLI first and who keeps using it. The outcomes study looked at how the number of merged PRs changed for engineers who adopted Copilot CLI or Claude Code. Its headline figure is a 24.0% increase in merged PRs per engineer per day (95% confidence interval 14.5–33.7%) over the 115 days after the January 5, 2026 adoption date. That figure is measured against a synthetic control built from engineers who did not use the tools, and a placebo test that set a fake adoption date of October 6, 2025 showed no effect: −1.1% (−10.6 to +8.6%). The increase held for four months, at 29.4% in February and 20.0% in March–April. [5]

Values from the paper's body text and its macro files. Changes in chance are changes in odds from a logistic regression, not changes in the rates themselves. Merged PRs count only PRs merged within 28 days of creation.

The caveats to attach when citing this result are spelled out in the paper itself. Merged PR count is a proxy for output: it favors small, frequent PRs and misses quality costs such as added complexity. The paper itself acknowledges "that a merged PR is not the same as the value it delivers," and its conclusion says "The pressing open question is now about quality." The figure counts PRs at one company, in one period (early 2026) and in Azure DevOps only, and the paper also discloses that right after the observation window closed, Microsoft cut off Claude Code licenses for most engineers and had them move to Copilot CLI. The authors write that Microsoft "sells AI tools" and "owns GitHub, the maker of Copilot CLI," and that "our proximity may have shaped our questions, design, and interpretation." [5]

There is also a point of comparison. The Cursor study by He et al. (MSR 2026), which the paper cites, reported that in open-source repositories the increase in commits and lines of code after adopting Cursor disappeared within two to three months. The Microsoft paper gives two reasons its own result held for four months: a difference in tool generation, and a difference in the unit of analysis (repository averages vs. a within-person comparison of the same engineer). The two studies did produce different results, and which one generalizes has not been settled. [5]

7. Five reasons the numbers differ

Apply all five reasons, and what remains is simple. As of mid-2026, most professional developers use some form of agent at least once a week, and the share using one daily is lower. People control how much they delegate, and trust has not kept up with use. Near-causal evidence that output increased has come from only one company, on one metric.

8. Checklist to apply today

• When you quote a number, state its five conditions with it. Sample (who), definition (what counts as an agent), frequency (weekly or daily), period and survey source. Quote only "90%," and it will clash with the next survey.
• Measure your own team with the same definitions. Weekly active users, daily users, the share of PRs created by agents and the share that people actually merged. Using JetBrains' "at least once a week" definition as is lets you compare your numbers with outside figures.
• Judge the effect by merged PRs and review load together. The Microsoft result measured only merged PR counts. More PRs also mean more review time. If you do not track the two side by side, you will miss a review backlog that looks like a productivity gain.
• Look at adoption alongside peer usage. At Microsoft, the strongest predictor of adoption was usage by peers in the same parent organization. But the paper cannot separate peer influence from the tendency of similar people to cluster together, so treat this finding as a metric to watch, not as a basis for a rollout strategy.
• People already comfortable with another AI tool may be less likely to stick. Engineers who used Copilot in the IDE did try the CLI tool, but fewer of them stayed with it. They need a separate reason to switch.
• Recheck user shares by tool every six months. The leader in January 2026 (Copilot, 29%) is not the leader in May–July (Claude Code, 39%). How the tools differ in config formats and approval defaults is covered in our comparison of five coding agents.
• Turn the trust numbers into a review policy. 63% of respondents rarely allow autonomous execution. A team policy that spells out what runs automatically and what needs approval is consistent with the survey results.

9. What this is not yet

• We did not run our own measurements. This article is a summary that cross-checks three survey announcements, one preprint and one vendor report.
• All of the surveys rely on voluntary responses. Weighting adjusts for region, experience and language, but the bias from people with a strong interest in AI tools responding more often remains. Actual usage may be lower than the published figures.
• The user shares by tool come from a single survey source. We could not find an independent survey besides JetBrains that measured the same question at the same time. Stack Overflow's 2026 annual survey opened on June 23, but its results had not been published at the time of writing (2026-09-17).
• JetBrains' three code-share values (47%, 38%, 27%) add up to more than 100%. We copied the announcement's averages as published; it was not disclosed whether the question allowed multiple answers or used range averages, so we did not adjust them.
• The Microsoft result covers one company, one period and one metric. All the authors work at Microsoft, and for the gap between the tools (Copilot CLI at about 2.2 times Claude Code), the authors themselves put forward organizational factors as a hypothesis. There is no quality metric.
• The peer effect is not causal. The paper states that it cannot separate peer influence from the tendency of similar people to cluster together.
• The Anthropic report has no methodology. It gives no sample size or period, so we cited only the direction (60% use, 0–20% fully delegated) and did not use its figures as a baseline in the comparison table.
• The pulse survey's 1,100 respondents are too small a sample to trust the absolute values. The 2025 annual survey value (31%) and the 2026 pulse value (59%) that Stack Overflow compared come from different sampling frames, so we read that comparison only for its direction.
• Figures cited in Korean news articles, such as "Claude Code at 75% in small teams" (our translation), could not be found in the original announcements, so we left them out.

10. Sources

All checked on 2026-09-17. For [5], we checked the body text and the numeric macro files of the arXiv LaTeX source directly. Survey announcements are the survey owners' own descriptions; each owner's stake is listed in the table in section 2.

Editorial standard: for every usage figure, we stated its sample, definition, frequency, period and survey source, and we put the scope of the output increase (one metric at one company) before the conclusions. We did not include figures from Korean news articles that we could not find in the original announcements.