ChatGPT 
Summary
AI coding statistics measure different things: how many developers use a tool, how much code it generates, how it affects delivery and how much users trust its output. Those measures cannot be treated as interchangeable.
In its May–July 2026 research, JetBrains reported weekly AI coding-agent use among 90% of professional developers and daily use among 68%. Separately, Stack Overflow’s 2025 survey reported 33% trust in the accuracy of AI output. The figures below keep each source, population and measurement period visible.
Key AI coding statistics for 2026
The highlights below distinguish tool adoption, reported output, delivery effects, trust and code security. Each figure is tied to the source’s own measure and population.
| Finding | Measure and scope | Source |
|---|---|---|
| 90% weekly; 68% daily | Professional developers using AI coding agents at work, including local and remote agents; research fielded May–July 2026. | JetBrains Developer Ecosystem Survey 2026 (more than 15,000 professional developers worldwide) |
| 84% use or plan to use AI; 51% use it daily | Stack Overflow 2025 survey respondents reported current use or plans for AI tools in development; the daily-use statistic is for professional developers. These broader categories are not equivalent to weekly/daily agent use. | Stack Overflow Developer Survey 2025 |
| 33% trusted AI output accuracy — Stack Overflow 2025 | Answers to the survey question about trust in the accuracy of AI output. This is reported confidence, not measured code correctness. | Stack Overflow Developer Survey 2025 |
AI coding tools adoption statistics
JetBrains asks about current AI-agent use at work, while Stack Overflow asks about current use or plans for future AI-tool use in development. The latter includes intended future use, so the responses are not equivalent and should not be combined into one adoption rate.
AI coding tool use at work
JetBrains reports how professional developers use the tools at work, not market share. The survey asked about multiple tools, so the results are not mutually exclusive.
| Tool | Reported work use | Comparison in JetBrains research |
|---|---|---|
| Claude Code | 39% in May–July 2026 | 18% in January 2026 |
| GitHub Copilot | 21% in May–July 2026 | 29% a year earlier |
| OpenAI Codex | 16% in May–July 2026 | 3% in January 2026 |
| Cursor | 12% in May–July 2026 | 18% in January 2026 |
Source: JetBrains, AI Coding Agents: Adoption Trends. The separate “main tool” answer is not shown here because it uses a different measure from reported use at work.
AI code generation statistics
AI-generated code, code written with AI assistance, accepted suggestions and total lines of code are different measures. A percentage from one cannot be substituted for another. Keep the population and counting method beside every reported share.
Google reported at Cloud Next in April 2026 that AI generated 75% of its new code, with engineers approving the output. This is a company-specific disclosure about new code at Google. It does not establish the AI-generated share of all software, all developers’ code or Google’s existing codebase.
At Meta’s April 2025 LlamaCon, Microsoft CEO Satya Nadella estimated that 20%–30% of code inside Microsoft repositories and some projects was written by software. He did not specify a single production-wide denominator or counting method. In the same discussion, Meta CEO Mark Zuckerberg said the company expected AI to do about half of development work within the following year; this was a forecast made in 2025, not a measured share of Meta’s code. Watch the official LlamaCon conversation.
A classifier-based study by Daniotti and colleagues estimated that AI generated about 29% of Python functions created or substantially changed by U.S. contributors to public GitHub projects by the end of 2024. The November 2025 version analyzed data from 2019–2024, extracting about five million functions from 31 million contributions. This is a model-based estimate for public GitHub Python functions, not a survey or a measure of all software. The authors’ earlier version used different data and reported a different estimate; this article uses the v2 finding only.
JetBrains asked more than 15,000 professional developers what share of the work code they produced in the previous month was fully generated by agents, partly AI-assisted or fully manual. The survey found that around 22% reported more than 80% of their code was agent-generated, while about one in five reported writing zero code without AI help. These are respondents’ estimates of their own work, not instrumented measurements of code repositories.
Want engineers who already work with AI agents?
Uvik Software’s senior engineers use AI coding agents in daily work, with review and testing rules.
Does AI make developers faster? What the research shows
Results vary with the task, developer, tool and study method. A controlled task experiment, a field experiment on familiar repositories and a self-reported productivity survey do not answer the same question.
METR’s early-2025 experiment found that participating experienced open-source developers took 19% longer on assigned tasks when AI tools were allowed. Its February 2026 follow-up said selection effects made newer results an unreliable estimate of the current productivity effect. Treat the earlier result as evidence from a specific experiment, not a forecast for every team or current tool.
In the 2025 study, 16 experienced developers working on repositories they knew contributed 246 real issues; each assigned task averaged about two hours. METR’s February 2026 update gives the original effect a 95% confidence interval of 2%–39% longer. Its later raw estimates were too weak to establish a precise current effect.
A different controlled experiment published in 2023 recruited 95 professional programmers in 2022 to implement a JavaScript HTTP server. Among developers who completed the task, the GitHub Copilot group finished 55.8% faster than the control group (95% confidence interval: 21%–89%). This result concerns one bounded programming task and should not be generalized to large codebases or end-to-end delivery.
A 2026 Management Science paper by Cui and colleagues pooled three randomized field experiments at Microsoft, Accenture and an unnamed Fortune 100 company, covering 4,867 developers granted or not granted a coding assistant. The pooled estimate was 26.08% more completed tasks (standard error 10.3%); individual experiments were noisy and the finding concerns task counts in those company settings, not a universal productivity or quality effect. Read the authors’ paper.
The same v2 working paper modeled quarterly output as the estimated share of AI-generated functions increased from zero to about 29% of US Python functions by the end of 2024. The model estimated a 3.6% increase in online code contributions over 2019–2024 associated with that change; the estimate was driven by experienced developers, with no statistically significant effect for inexperienced developers. This modeled contribution-volume estimate is distinct from generated-code share and is not a direct measure of work hours or task speed. Read the paper.
Stack Overflow’s 2025 survey offers a separate perception measure: 52% of developers said AI tools or agents had a positive effect on their productivity in the past year. A perceived productivity effect is not a measured change in delivery time, quality or team throughput.
Do developers trust AI-generated code?
In Stack Overflow’s 2025 survey, 33% of respondents to the accuracy question trusted AI output, while 46% distrusted it. Only 3% reported high trust. This measures reported confidence in accuracy, not the proportion of generated code that is correct.
On a separate question, 66% of respondents said one frustration was AI solutions that were “almost right, but not quite”; 45% said debugging AI-generated code was more time-consuming. These are self-reported experiences from a multi-select survey question, not measured rework rates for delivered software. The survey also found that 29% of professional developers said AI tools handled complex tasks poorly; that figure concerns task performance, not trust in accuracy.
Security and code quality
Security benchmarks can reveal failure modes, but their task-level results are not production incident rates. Their numbers depend on which models, prompts, languages and vulnerability classes were tested.
Veracode’s 2025 GenAI Code Security Report tested more than 100 language models on 80 deliberately security-sensitive function-completion tasks across Java, JavaScript, C# and Python. Across those model-task completions, 45% introduced a detectable security flaw. The benchmark used intentionally minimal prompts without security-specific guidance and checked generated snippets using static analysis. This is a controlled benchmark failure rate, not a claim that 45% of production AI code is vulnerable.
A paper presented at USENIX Security 2025 tested 16 models using two prompt datasets, generating 576,000 Python and JavaScript samples. It reported hallucinated package references averaging at least 5.2% for commercial models and 21.7% for open-source models; the denominator is package references, not generated samples. The study also identified 205,474 unique nonexistent package names. Those names are not all known malicious packages, and the experiment does not measure production prevalence.
Earlier research by Pearce and colleagues, submitted in 2021 and accepted for IEEE Security and Privacy 2022, found about 40% of 1,689 programs generated by GitHub Copilot across 89 high-risk weakness scenarios were vulnerable. This is a historical Copilot-era benchmark, not a measurement of current coding assistants or deployed code.
AI coding tool market statistics and forecasts
Market signals use different denominators. GitHub’s 2025 Octoverse report said more than 1.1 million public repositories used an LLM SDK; 693,867 of those projects were created in the preceding 12 months, a reported 178% year-over-year increase (August 2025 versus August 2024). This measures public repositories using an LLM SDK, not code written by AI. GitHub Octoverse 2025.
Anthropic reported in February 2026 that Claude Code’s run-rate revenue had exceeded $2.5 billion and more than doubled since the beginning of 2026. Run-rate revenue is a company-reported annualized pace, not recognized revenue for a completed year or a measure of the overall AI coding market. Anthropic’s February 2026 announcement.
Microsoft reported on its January 2026 earnings call that GitHub Copilot had more than 4.7 million paid subscribers, up 75% year over year. This is a company-reported paid-subscriber count, not the share of developers using the tool. Microsoft FY2026 Q2 earnings call.
Adoption forecasts describe expected future use, not measured outcomes. Gartner’s July 2025 press release forecast that 90% of enterprise software engineers would use AI code assistants by 2028, from fewer than 14% in early 2024. The public release does not provide the survey method behind those estimates. Gartner’s August 2025 forecast that 40% of enterprise applications, up from fewer than 5% in 2025, would integrate task-specific AI agents by the end of 2026 used a different denominator and should not be compared with the share of developers using assistants.
What these statistics mean for engineering leaders
These findings answer different questions: whether developers use tools, how much work they estimate is AI-generated, how a specific experiment changed completion time, what people say they trust, and how generated snippets perform in security tests. None is a substitute for measuring a team’s own delivery cycle.
- Define whether a measure is current use, intended use, generated output, assisted output, perceived productivity or measured task time.
- Record the population, question, period, task and denominator before comparing results.
- Measure the full delivery cycle, including review, testing, debugging and rework, alongside implementation time.
- Test generated code for correctness, security and dependency validity before accepting it.
- Use a controlled internal comparison when making decisions about a particular codebase or team.
For related guidance, see AI-augmented software development, AI staff augmentation, AI-powered software development and risks of AI in software development. For hiring information, see hire AI and ML engineers.
Add senior AI engineers to your team
Matched profiles within 48 hours after a signed SOW. $50 to $99 per hour. 30-day no-cost replacement.
Methodology and sources
This article compiles findings from primary research and company disclosures. Each statistic keeps the source date, population and metric definition. Survey adoption, generated-code share and measured productivity describe different outcomes. We do not combine them into a single industry estimate or infer a trend from studies with different methods. Company disclosures describe the reporting company unless the source states otherwise.
- Meta Developers: LlamaCon 2025 conversation with Microsoft CEO Satya Nadella
- Anthropic: Claude Code run-rate revenue disclosure, February 2026
- Microsoft: GitHub Copilot subscriber disclosure, January 2026
- GitHub Octoverse 2025
- Cui et al.: three company field experiments, Management Science (2026); authors’ preprint
- JetBrains: AI coding-agent adoption trends, 2026
- JetBrains: how much code developers report agents write, 2026
- Stack Overflow Developer Survey 2025: AI
- Google Cloud Next 2026 remarks
- Daniotti et al., Who is using AI to code? (version 2, 2025)
- METR: early-2025 developer productivity experiment
- METR: February 2026 experiment update
- Peng et al.: GitHub Copilot controlled task experiment (2023)
- Veracode: 2025 GenAI Code Security Report
- USENIX Security 2025: package hallucinations in code-generating models
- Pearce et al.: Copilot security study (2021–2022)
- Gartner: software engineering trends and forecasts, July 2025
- Gartner: enterprise-application agent forecast, August 2025
- DORA 2025 report
Related reading: AI engineer salary, Python developer statistics and whether AI will replace programmers. For questions about an engineering workflow, talk to us.
Updated for 2026. Survey periods and study dates above refer to the sources’ original evidence windows.
How to cite this page
Uvik Software (2026). AI Coding Assistant and Code Generation Statistics 2026. https://uvik.net/blog/ai-coding-assistant-statistics/
Further reading: best AI technology podcasts in 2026 and the best agentic AI frameworks.
