Developer surveys now agree on one thing: most professional developers use AI at work. JetBrains found 90% using AI coding agents at least weekly in May–July 2026, and Stack Overflow’s 2026 survey reports that most respondents use coding assistants and agents every day. They disagree on almost everything else, because each asks a different question of a different population.
This page lines up the latest editions of the main sources: Stack Overflow’s 2026 Developer Survey (published October 2026), JetBrains’ Developer Ecosystem Survey 2026 (August 2026), GitHub’s Octoverse 2026 (October 28, 2026), the DX reports for Q4 2025 and Q2 2026, and Google’s 2025 DORA report. It adds the controlled studies on productivity and the measured work on code quality and security. Every figure sits in a table row with its type, its population and its source, and the older figures that circulate widely are corrected in the “Often quoted as” section.
Digital Applied did not run any of these surveys. We read each one at the publisher, state its sample, and keep survey claims apart from vendor claims and randomised studies. For tool selection and costs, see our AI coding IDE landscape and per-tool cost calculator; for the money side of the same question, see AI agent productivity statistics.
Each row gives the figure, what it measures (population, metric, period) and a type. Survey means self-reported answers from a sample. Benchmark (platform data) is telemetry from a vendor’s own customers, so it is not a random sample. Benchmark (vendor-run) is a test the vendor designed and ran. Official (company-reported) is a company’s own count of its platform. Measured (third-party) is an outside analysis of code. Academic study covers randomised trials and papers. Compare figures only inside one source and one period.
- 01Use is near-universal, and agents are now the headline.JetBrains reports 90% of professional developers using AI coding agents at work at least weekly and 68% daily (May–July 2026, more than 15,000 respondents). Stack Overflow 2026 reports 73% of respondents using coding assistants and agents daily. DORA 2025 and DX report 90% or more on their own populations.
- 02Tool leadership changed within a year.In JetBrains’ data, Claude Code work adoption went from 18% in January 2026 to about 39% in May–July 2026, Codex from 3% to 16%, while GitHub Copilot fell from 29% to 21% and Cursor from 18% to 12%. Copilot is still one of the most widely known tools, at 79% awareness.
- 03Agents are changing how code reaches GitHub.GitHub’s Octoverse 2026 says more than 40% of pull requests started with an agent’s first draft by September 2026, up from 17% in March, and that 55% of Copilot monthly active users now use agentic features.
- 04Time saved is real, but value is harder to see.DX measured 3.6 hours saved per week in Q4 2025 and over 6 hours for heavy users in Q2 2026, yet DX reports median quarterly AI spend up nearly 28x in the Tech sector alone over a year, with the innovation ratio barely moving. METR’s randomised study found experienced developers 19% slower in early 2025, and its later follow-up could not produce a reliable estimate.
- 05Trust is conditional and quality signals are mixed.In Stack Overflow 2026, 48% trust AI output when they can easily verify it and 6.6% trust it for important work decisions. Veracode’s test found models chose the insecure option 45% of the time, and DX saw change confidence fall 6.1% while maintainability rose 3.8%.
01 — AdoptionAdoption rates: roughly nine in ten, depending on who is asked.
The adoption rows below use five different questions. Stack Overflow asks all respondents, including students and hobbyists, how they use AI. JetBrains weights its sample to the global population of professional developers and asks about work use. DORA asks technology professionals whether they use AI at work. DX measures developers inside organisations that already use its platform, which skews toward companies that invest in developer tooling.
Two cautions apply before comparing any of them. First, a survey of everyone who answers a public developer questionnaire over-represents people who are online and engaged with developer communities, and Stack Overflow’s own sample includes learners and career changers. Second, a weekly-use question and a daily-use question measure different habits, so a drop from 90% to 68% inside the JetBrains data is a change of question, not a change of behaviour.
Stack Overflow’s 2026 edition drew more than 30,000 respondents from 169 countries. Its figures replace the 2025 numbers that were widely quoted last year, but the 2025 rows stay in the table because DORA and most vendor reports still cite them.
| Figure | What it measures | Type | Source |
|---|---|---|---|
| 84% | Share of Stack Overflow 2025 respondents using or planning to use AI tools in their development process (76% in 2024); superseded by the 2026 edition (fielded mid-2025) | Survey | Stack Overflow, 2025 Developer Survey |
| 51% | Share of professional developers using AI tools daily, Stack Overflow 2025 survey (fielded mid-2025) | Survey | Stack Overflow, 2025 Developer Survey |
| 73% | Share of Stack Overflow 2026 respondents who use coding assistants and agents daily (base is respondents who use them; the page states no n for this figure) (2026 survey, published Oct 2026) | Survey | Stack Overflow, 2026 Developer Survey |
| 66% | Share of Stack Overflow 2026 respondents naming coding assistants and agents as a use of AI, ahead of general chatbots at 63% (survey of more than 30,000 technologists in 169 countries) (2026 survey) | Survey | Stack Overflow, 2026 Developer Survey |
| 31% | Share of daily AI users who use it more than four hours a day, Stack Overflow 2026 (over 70% of respondents use AI daily; n = 10,926 for the time breakdown) (2026 survey) | Survey | Stack Overflow, 2026 Developer Survey |
| 90% | Share of professional developers using AI coding agents at work at least weekly (68% daily), JetBrains Developer Ecosystem Survey 2026, more than 15,000 professional developers, reweighted to the global developer population (May–Jul 2026) | Survey | JetBrains, Developer Ecosystem Survey 2026 |
| 90% | Share of developers regularly using at least one AI tool at work for coding in January 2026 (74% used specialised AI developer tools), JetBrains AI Pulse, more than 10,000 professional developers (Jan 2026) | Survey | JetBrains, AI Pulse (January 2026) |
| 90% | Share of survey respondents reporting that they use AI at work, DORA 2025 (nearly 5,000 technology professionals worldwide) (2025) | Survey | Google Cloud, 2025 DORA Report |
| 90%+ | Share of developers using AI tools monthly or more, DX Q4 2025 report (135,000 developers across 425 organisations on DX’s platform, so not a random sample) (Q4 2025) | Benchmark (platform data) | DX, Q4 2025 report |
| 90%+ | Industry-wide AI utilisation reported in DX’s Q2 2026 analysis of more than 500 engineering organisations (Apr–Jun 2026) | Benchmark (platform data) | DX, Q2 2026 report |
| 62% | Share of Stack Overflow 2026 respondents with a very or somewhat favourable attitude towards AI tools in their work (more than 60%, per the survey home page) (2026 survey) | Survey | Stack Overflow, 2026 Developer Survey |
02 — ToolsTool preferences: Claude Code and Codex rose while Copilot and Cursor slipped.
JetBrains runs the most useful series for tool comparison because it asks about work use and repeats the question. Its AI Pulse survey in January 2026 covered more than 10,000 professional developers; its Developer Ecosystem Survey 2026 covered more than 15,000 in May–July 2026. JetBrains states that its AI Pulse survey was promoted through Instagram ads and a research panel, with no mention of AI in the survey promotion, which is meant to avoid attracting only enthusiasts or sceptics. Both are reweighted, but they are separate samples, so small changes should be read as direction rather than precise movement.
Stack Overflow’s 2026 survey reports Claude Code and GitHub Copilot as the leading coding agents (66% and 59%), but the page does not state the base for that figure, so it should not be read as a share of all developers. The JetBrains rows are the better guide to market share.
| Figure | What it measures | Type | Source |
|---|---|---|---|
| 39% | Share of professional developers using Claude Code at work (47% in the United States; 18% in January 2026), JetBrains Developer Ecosystem Survey 2026 (May–Jul 2026) | Survey | JetBrains, Developer Ecosystem Survey 2026 |
| 31% | Share of developers for whom Claude Code is the single most used AI coding tool, JetBrains 2026 (May–Jul 2026) | Survey | JetBrains, Developer Ecosystem Survey 2026 |
| 21% | Share of professional developers using GitHub Copilot at work (29% in January 2026; 79% awareness), JetBrains 2026 (May–Jul 2026) | Survey | JetBrains, Developer Ecosystem Survey 2026 |
| 16% | Share of professional developers using Codex at work (3% in January 2026; awareness 27% to 65%), JetBrains 2026 (May–Jul 2026) | Survey | JetBrains, Developer Ecosystem Survey 2026 |
| 12% | Share of professional developers using Cursor at work (18% in January 2026; awareness 75%), JetBrains 2026 (May–Jul 2026) | Survey | JetBrains, Developer Ecosystem Survey 2026 |
| 7% / 6% / 9% | Work adoption of OpenCode (7%), Google Antigravity (6%) and JetBrains AI in IDEs or Junie (about 9%) among professional developers, JetBrains 2026 (May–Jul 2026) | Survey | JetBrains, Developer Ecosystem Survey 2026 |
| 66% / 59% | Claude Code and GitHub Copilot as the leading coding agents in Stack Overflow 2026 (the page does not state the respondent base; read as the share of agent users) (2026 survey) | Survey | Stack Overflow, 2026 Developer Survey |
| 29% / 40% | GitHub Copilot work adoption among developers globally (76% awareness) and at companies with over 5,000 employees, JetBrains AI Pulse (Jan 2026) | Survey | JetBrains, AI Pulse (January 2026) |
| 91% / 54 | Claude Code satisfaction (CSAT) and Net Promoter Score among its users, JetBrains AI Pulse (Jan 2026) | Survey | JetBrains, AI Pulse (January 2026) |
| 28% | Share of developers using the ChatGPT chatbot for coding tasks at work (Gemini 8%, Claude chatbot 7%), JetBrains AI Pulse (Jan 2026) | Survey | JetBrains, AI Pulse (January 2026) |
03 — DivergenceWhere the surveys disagree, and why.
The same metric often has several values because the surveys differ in population, wording and date. The table sets the closest comparable measures side by side. A gap between two cells usually reflects the question asked, not an error in either survey. The practical rule is to quote a figure with its survey name, its year and its population every time, and to avoid putting two surveys’ numbers in the same chart without saying so.
| Metric | Stack Overflow | JetBrains | DORA | DX |
|---|---|---|---|---|
| Any AI use | Stack Overflow 2025: 84% use or plan to (76% in 2024) | JetBrains Jan 2026: 90% regularly use a tool at work | DORA 2025: 90% use AI at work | DX Q4 2025: 90%+ monthly |
| Daily use | Stack Overflow 2025: 51% of professionals use AI tools daily; 2026: 73% use coding assistants and agents daily | JetBrains May–Jul 2026: 68% use coding agents daily | Not reported | Not reported |
| GitHub Copilot at work | Not asked on the same basis | JetBrains: 29% (Jan 2026), 21% (May–Jul 2026) | Not reported | Not reported |
| Trust | Stack Overflow 2025: 33% trust accuracy, 46% distrust | Stack Overflow 2026: 48% trust output they can easily verify | DORA 2025: 30% little or no trust in AI code | Not reported |
| Share of code written by AI | Not reported | Not reported | Not reported | DX: 22% of merged code (Q4 2025); 52.7% of all code on a new measure (Q2 2026) |
04 — AgentsAgents on GitHub: first drafts by agents are now routine.
GitHub’s Octoverse 2026 covers the 12 months to September 30, 2026 across public and private repositories. These are GitHub’s own platform counts, so they describe activity on GitHub and not the whole industry. They are the most direct measure available of how much work agents now start.
The report also notes that agent-started pull requests change multiple files more often (87% against 79%) and merge within two weeks at a higher rate by September (90%). GitHub says review is now how developers shape what an agent returns. Our 30-60-90 day rollout plan covers how teams set review and permissions around this.
| Figure | What it measures | Type | Source |
|---|---|---|---|
| 40%+ | Share of pull requests on GitHub that started with an agent’s first draft in September 2026 (17% in March 2026; 110 million pull requests in September), Octoverse 2026 (Sep 2026) | Official (company-reported) | GitHub, Octoverse 2026 |
| 55% | Share of GitHub Copilot monthly active users using agentic features (35% a year earlier), Octoverse 2026 (Sep 2026) | Official (company-reported) | GitHub, Octoverse 2026 |
| 90% | Share of agent-involved pull requests merged within two weeks in September 2026 (83% in February; about 86% for pull requests without detected agent involvement) (Feb–Sep 2026) | Official (company-reported) | GitHub, Octoverse 2026 |
| 87% vs 79% | Share of agent-started pull requests that change multiple files, against pull requests without detected agent involvement (2026) | Official (company-reported) | GitHub, Octoverse 2026 |
| 17.9 billion | Pushes to GitHub in the 12 months to September 30, 2026, more than 2024 and 2025 combined (commits 34.3 billion; merged pull requests 1.3 billion) (Oct 2025–Sep 2026) | Official (company-reported) | GitHub, Octoverse 2026 |
| 225 million+ | Developers served by GitHub (United States 38.3 million, India 32 million) (Sep 2026) | Official (company-reported) | GitHub, Octoverse 2026 |
| 44.1% / 41.1% / 40.4% / 21.2% | Share of merged pull requests with detected AI involvement: TypeScript, Python, Rust and Java (12 months to Sep 2026) | Official (company-reported) | GitHub, Octoverse 2026 |
05 — ProductivityProductivity evidence: self-reports, platform data and randomised studies.
Three kinds of evidence answer the productivity question, and they point in different directions. Surveys ask developers how they feel. DX telemetry measures pull requests and survey responses inside its customers. Randomised studies assign AI at random, which is the only design that can separate cause from selection.
The randomised results are mixed and dated. A pooled analysis of three field experiments with 4,867 developers found 26.08% more completed tasks. METR’s 2025 trial found experienced open-source developers 19% slower. METR’s later follow-up says its own data is only very weak evidence, because developers who most value AI declined to take part and skipped tasks. Treat any hours-saved figure built on these studies as an estimate to test on your own team, not a planning constant.
| Figure | What it measures | Type | Source |
|---|---|---|---|
| 3.6 hours | Average weekly time saved per developer using AI coding assistants (gains have plateaued), DX Q4 2025 (135,000 developers, 425 organisations) (Q4 2025) | Benchmark (platform data) | DX, Q4 2025 report |
| 22% | Share of merged code that is AI-authored, DX Q4 2025 (Q4 2025) | Benchmark (platform data) | DX, Q4 2025 report |
| 60% | Extra pull requests shipped by daily AI users against non-users, DX Q4 2025; a correlation, with mixed results on quality (Q4 2025) | Benchmark (platform data) | DX, Q4 2025 report |
| 6+ hours | Weekly time saved by heavy AI users in DX’s Q2 2026 report (500+ organisations); DX says the saved hours are not yet visibly converting into new value (Apr–Jun 2026) | Benchmark (platform data) | DX, Q2 2026 report |
| 1.42 to 1.94 | Median weekly pull requests per engineer (+37% over four quarters), DX Q2 2026; the Developer Experience Index fell from 67 to 65 over the same window (Q3 2025–Q2 2026) | Benchmark (platform data) | DX, Q2 2026 report |
| 80%+ | Share of respondents who believe AI has increased their productivity, DORA 2025 (2025) | Survey | Google Cloud, 2025 DORA Report |
| 52% | Share of Stack Overflow 2025 respondents agreeing AI tools or agents had a positive effect on their productivity (2025 survey) | Survey | Stack Overflow, 2025 Developer Survey |
| +26.08% | Increase in completed tasks for developers given an AI coding assistant, pooled across three randomised field experiments (4,867 developers at Microsoft, Accenture and a Fortune 100 company) (trials reported 2025) | Academic study | Microsoft Research, three field experiments |
| +19% | Change in task completion time when 16 experienced open-source developers were allowed to use early-2025 AI tools (246 tasks; AI made them slower) (Feb–Jun 2025) | Academic study | METR, arXiv 2507.09089 |
| −18% / −4% | Estimated speedup in METR’s later study: returning developers −18% (interval −38% to +9%), new recruits −4% (−15% to +9%); METR calls it weak evidence because of selection effects (Aug–late 2025) | Academic study | METR, experiment update (Feb 2026) |
| 30%–50% | Share of developers in METR’s later study who left out some tasks because they did not want to do them without AI (late 2025) | Academic study | METR, experiment update (Feb 2026) |
06 — QualityCode quality: more code, less reuse.
GitClear’s analysis of 211 million changed lines (2020–2024) is the most cited measurement of how code looks after AI assistants arrived. It reports copy/pasted lines rising from 8.3% to 12.3% of changed lines and moved, or refactored, lines falling from 25% to under 10%. GitClear sells code analytics, and its link between these trends and AI assistants is its own interpretation of correlated data.
DX’s Q2 2026 report adds a platform view of the same tension: maintainability improved while developers’ confidence that changes will not break something fell. Stack Overflow 2025 respondents named almost-right answers (66%) and slower debugging (45%) as their main frustrations. Our vibe coding anti-patterns guide covers the habits that produce these results.
| Figure | What it measures | Type | Source |
|---|---|---|---|
| 12.3% | Share of changed lines classed as copy/pasted in 2024, up from 8.3% in 2021; 211 million changed lines, 2020–2024 (2020–2024) | Measured (third-party) | GitClear, AI Copilot Code Quality 2025 |
| 25% to under 10% | Share of changed lines classed as moved (refactoring) in 2021 against 2024, GitClear (2021–2024) | Measured (third-party) | GitClear, AI Copilot Code Quality 2025 |
| +3.8% / −6.1% | Change in code maintainability and in change confidence over four quarters, DX Q2 2026 (Q3 2025–Q2 2026) | Benchmark (platform data) | DX, Q2 2026 report |
| 52.7% | Share of all code that is AI-generated in DX’s Q2 2026 measure (24% two quarters earlier); DX defines this differently from the 22% in its Q4 2025 report (Apr–Jun 2026) | Benchmark (platform data) | DX, Q2 2026 report |
| 66% / 45% | Developers naming “AI solutions that are almost right, but not quite” (66%) and “debugging AI-generated code is more time-consuming” (45%) as frustrations, Stack Overflow 2025 (2025 survey) | Survey | Stack Overflow, 2025 Developer Survey |
07 — TrustTrust and security: verify first.
Developers use AI heavily but trust it narrowly. In Stack Overflow’s 2026 survey the largest group (48%) trusts AI output when they can easily verify it, and 6.6% trust it for important work decisions. The security evidence in this list is a single vendor benchmark, so treat it as a signal about model behaviour on one test set and not as the rate for production code.
Veracode designed 80 tasks around known weakness types and ran more than 100 models. In 45% of cases the model chose the insecure method over the secure one, and Java failed over 70% of the time. Teams that need controls around this should read our AI transformation services overview and keep scanning rules identical for human and AI-written code.
| Figure | What it measures | Type | Source |
|---|---|---|---|
| 45% | Share of test cases in which models chose the insecure method when offered a secure and an insecure one, Veracode (80 tasks, more than 100 LLMs) (2025) | Benchmark (vendor-run) | Veracode, 2025 GenAI Code Security Report |
| 70%+ | Security failure rate for Java, the riskiest language in Veracode’s test; Python, C# and JavaScript between 38% and 45% (2025) | Benchmark (vendor-run) | Veracode, 2025 GenAI Code Security Report |
| Figure | What it measures | Type | Source |
|---|---|---|---|
| 48.0% | Share of Stack Overflow 2026 respondents who trust AI output when they can easily verify it (2026 survey) | Survey | Stack Overflow, 2026 Developer Survey |
| 6.6% | Share of Stack Overflow 2026 respondents who trust AI for many tasks including important work decisions (2026 survey) | Survey | Stack Overflow, 2026 Developer Survey |
| 12.3% | Share of Stack Overflow 2026 respondents who do not trust AI for most work tasks (2026 survey) | Survey | Stack Overflow, 2026 Developer Survey |
| 46% vs 33% | Share of Stack Overflow 2025 respondents who distrust the accuracy of AI tools against those who trust it (3% highly trust) (2025 survey) | Survey | Stack Overflow, 2025 Developer Survey |
| 30% | Share of DORA 2025 respondents with little or no trust in AI-generated code (2025) | Survey | Google Cloud, 2025 DORA Report |
08 — ChecksFigures often quoted that the sources do not support.
Several figures from earlier versions of this page, and from other roundups, do not match the publisher’s own page. The corrections below come from re-reading each source on November 5, 2026.
| Often quoted as | What the source says |
|---|---|
| Stack Overflow: trust in AI accuracy fell to 29%, down from 40% | The 2025 survey reports 33% trust the accuracy of AI tools and 46% distrust it. The 29% figure in that survey is the share of professionals who believe AI tools struggle with complex tasks (35% in 2024). |
| DX: 85,350 developers at 435 companies | DX’s Q4 2025 report states 135,000 developers across 425 organisations. Its 2.3 versus 1.4 pull requests per week split is not on the report page; the page states daily users ship 60% more PRs. |
| Veracode: 45% of AI-generated code contains a vulnerability; Java 72% | Veracode’s release says models chose the insecure option 45 percent of the time when given a choice, and that Java’s failure rate was over 70 percent. |
| Claude Code 18%, Cursor 18%, Copilot 29% at work | Those are JetBrains January 2026 figures. By May–July 2026 the same survey series shows about 39%, 12% and 21%. |
| Stack Overflow 2025: 84% adoption is the latest | Stack Overflow published its 2026 survey on October 6, 2026 with more than 30,000 respondents; use that edition and name the year. |
| Apiiro, McKinsey and Microsoft Work Trend figures on earlier versions | These were removed because they were read from secondary coverage or did not measure software development. They can be cited only from the publisher’s own page. |
09 — The SourcesSources, grouped by publisher.
Every figure above was read on the publisher’s own page between October and November 5, 2026. Publications are grouped by the type of publisher, with the publication date or the date read.
Developer surveys and research programmes
- Stack Overflow. 2026 Developer Survey (published Oct 6, 2026) and 2025 Developer Survey.
- JetBrains Research. Developer Ecosystem Survey 2026 adoption post (Aug 2026) and AI Pulse post (Apr 2, 2026).
- Google Cloud, DORA research programme. Announcing the 2025 DORA Report, Sep 23, 2025.
Platform and company reports
- GitHub. Octoverse: Commits, pushes reach new highs in 2026 as developers build with agents, Oct 28, 2026.
- DX. AI-Assisted Engineering Q4 Impact Report (read Nov 5, 2026) and State of AI Impact in Engineering Q2 report announcement, Nov 4, 2026.
- GitClear. AI Copilot Code Quality: 2025 Look Back at 12 Months of Data (read Nov 5, 2026).
- Veracode. 2025 GenAI Code Security Report press release, Jul 30, 2025.
Academic and research-lab studies
- METR. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, arXiv, Jul 12, 2025.
- METR. We are Changing our Developer Productivity Experiment Design, Feb 24, 2026.
- Microsoft Research. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers, Jun 2025.
Near-universal use, uneven proof of value.
Across the sources, use is no longer the open question: roughly nine in ten professional developers use AI at work, and the leading tools changed within a year. What the data does not yet show is where the time saved ends up. DX’s telemetry shows output rising while its Developer Experience Index fell from 67 to 65 and the innovation ratio barely moved, while the best randomised evidence dates from 2025 or is self-limited.
For teams planning the next half-year, measure daily use and pull-request outcomes rather than licence counts, keep review and scanning rules the same for AI-written code, and cite each survey with its year and population. The productivity statistics roundup and our AI transformation services cover the next step.