Professional developers say AI agents fully wrote about 47% of the code they produced for work in the past month. That average comes from JetBrains’ Developer Ecosystem Survey 2026, fielded May–July 2026 across more than 15,000 professional developers.
Below are verified vibe coding statistics on adoption, tools, regions and code security, each carrying its source and measurement period.
Vibe Coding Statistics
How Much Code Do Vibe Coding Agents Write?
JetBrains asked more than 15,000 professional developers to split last month’s output three ways: fully generated by AI agents, written with some AI assistance, or written manually.
The self-reported averages landed at roughly 47%, 38% and 27%. JetBrains notes these use bucket midpoints, so the three can add to more than 100%.
Over half of developers now write less than 20% of their code by hand. One in five writes no code at all without AI help.
Full hand-off remains a minority position. Around 22% report that agents generate over 80% of their code, and about a quarter of senior developers fall in that group.
Agent Share By Most-Used Tool
| Most-used AI coding tool | Share of its users generating over 80% of code with agents |
|---|---|
| OpenAI Codex | 42% |
| Claude Code | 32% |
| Cursor | 28% |
Source: JetBrains Developer Ecosystem Survey 2026, May–July 2026
Codex users skew heaviest: 37% of them report writing no code without AI assistance. Cursor users average 58% agent-generated code.
Tech stack splits the field. Developers naming Go, JavaScript or TypeScript as their main language report the highest agent-generated averages at 54%–55%, while Java and Python sit at 48%–51%.
C and C++ developers stay closest to the keyboard, reporting 38% of their code as manually written on average. Anyone still hand-writing front-end code can shortcut the boilerplate with a CSS property generator or an HTML element generator.
Vibe Coding Adoption Among Professional Developers
Stack Overflow put the term itself to developers in its 2025 survey, defining vibe coding as generating software from LLM prompts. The question drew 26,564 responses out of 49,000-plus participants across 177 countries.
Most said no: 72.2% answered “No” and a further 5.3% answered “No, emphatically”. Stack Overflow released the results on 29 December 2025.
Broad AI tool use runs far ahead of vibe coding specifically. In the same survey, 84% of respondents were using or planning to use AI tools, and 50.6% of professional developers used them daily.
Google’s DORA team found a similar split. Its 2025 State of AI-assisted Software Development report, based on nearly 5,000 technology professionals, recorded 90% using AI at work and a median of two hours per day spent working with it.
Trust lagged adoption in both datasets. DORA reported 24% trusting AI output a great deal or a lot, against 30% trusting it a little or not at all.
Vibe Coding Tools And Adoption Rates In 2026
JetBrains tracks tool adoption at work as a multi-select question, so developers can report several tools. Claude Code led at 39% globally in May–July 2026, rising to 47% in the United States.
GitHub Copilot slipped to 21% from 29% a year earlier, though it retains 79% awareness among developers. OpenCode reached 7% adoption with 42% mindshare and no large vendor behind it.
Movement over four months was sharp for three tools measured in both waves.
Claude Code doubled from 18% in January 2026, and Codex went from 3% to 16%. Cursor moved the other way, falling to 12% from 18%.
Codex awareness climbed from 27% to 65% across the same window. Cursor gained mindshare over those months, from 69% to 75%, while its work adoption dropped.
AI Agent Use At Work By Industry
Stack Overflow ran a pulse survey of 1,100 developers and working professionals in late April 2026. It found 59% using AI agents at work at any frequency, compared with 31% in its 2025 Developer Survey.
Oversight stayed tight. Some 63% rarely or never let agents run entirely on autopilot, and 60% block agents from making unapproved system changes.
Concerns eased but did not disappear. Respondents strongly agreeing that accuracy is a barrier fell to 47% from 57%, and those strongly agreeing on security concerns fell to 44% from 56%.
Who Is Vibe Coding Beyond Software Engineering?
Lovable published its first Build Economy report on 9 June 2026, drawing on platform usage data from January 2025 to May 2026 plus a May 2026 survey of its own users.
Lovable reports that 80% of people building on the platform are in self-identified non-technical roles, and that nearly two-thirds come from outside the technology industry.
Apps built there average 720 million visits per month, and 8 in 10 users intend to monetise what they build. These are the company’s own platform figures, not independently audited.
Formal training scaled too. Google and Kaggle’s 5-Day AI Agents: Intensive Vibe Coding Course drew more than 353,000 registered participants and over 6,000 capstone submissions, per Google’s 3 August 2026 recap.
Collins Dictionary named “vibe coding” its Word of the Year on 6 November 2025, citing a rise in usage since the term first appeared in February 2025. Much of what this cohort ships is web output, where a meta tag generator and clean FAQ structured data still decide how the page performs.
Vibe Coding Security Statistics
Veracode tested more than 100 large language models across four snapshots for its 2026 GenAI Code Security Report, published 28 July 2026.
The average security pass rate was 56%, barely moved from 55% in the first edition. Roughly 44% of code generation tasks introduced a known vulnerability.
Syntax is a solved problem by comparison: models produce syntactically correct code close to 100% of the time.
Pass rates split sharply by weakness type, from 87% on cryptographic algorithms down to 15% on cross-site scripting and 12% on log injection.
Security Pass Rate By Model Category
| Model grouping | Average security pass rate |
|---|---|
| GPT-5.5 (highest single model) | 68% |
| Reasoning models | 56% |
| Large models | 53% |
| General-purpose models | 52% |
| Coding-specialised models | 51% |
| Non-reasoning models | 51% |
| Medium models | 51% |
| Small models | 51% |
Source: Veracode 2026 GenAI Code Security Report, published July 2026. Groupings overlap; each row is an average pass rate, not a share of one total.
More than half of the models tested clustered at 50%–53%. Coding-specialised models did not outperform general-purpose ones.
Where These Vibe Coding Statistics Come From
| Source | Sample | Field or release period |
|---|---|---|
| JetBrains Developer Ecosystem Survey 2026 | 15,000+ professional developers | May–July 2026 |
| Stack Overflow Developer Survey 2025 | 49,000+ respondents, 177 countries | Released 29 Dec 2025 |
| Stack Overflow agent pulse survey | 1,100 developers and professionals | Late April 2026 |
| DORA State of AI-assisted Software Development | Nearly 5,000 technology professionals | Published 23 Sep 2025 |
| Veracode GenAI Code Security Report 2026 | 100+ models, four test snapshots | Published 28 Jul 2026 |
| Lovable Build Economy report | Platform usage plus user survey | Jan 2025–May 2026 |
Source: publisher documentation for each named survey and report, periods as listed
All figures above are opt-in or self-reported except the Veracode model tests, which are lab measurements. For wider context on the tooling side, see these web coding statistics for 2026.
FAQs
What percentage of code do AI agents write in 2026?
Professional developers report an average of about 47% of their code as fully agent-generated, per JetBrains’ Developer Ecosystem Survey 2026 covering May–July 2026. Around 22% say agents produce over 80% of their code.
Do most professional developers vibe code?
No. In Stack Overflow’s 2025 Developer Survey, 72.2% answered “No” and a further 5.3% answered “No, emphatically” when asked if vibe coding was part of their professional work.
Reddit threads claim people write zero code by hand. Is that accurate?
Partly. JetBrains found one in five professional developers writes no code without AI, but the average share of fully agent-generated code was about 47% in May–July 2026, not 100%.
Is Cursor losing ground, as Reddit users suggest?
JetBrains recorded Cursor at 12% work adoption in May–July 2026, down from 18% in January 2026. Claude Code reached 39% and GitHub Copilot fell to 21% over comparable periods.
How secure is AI-generated code?
Veracode’s 2026 report put the average security pass rate at 56% across 100-plus models, meaning roughly 44% of tested generation tasks introduced a known vulnerability. Cross-site scripting passed just 15% of the time.
Sources
https://blog.jetbrains.com/research/2026/08/how-much-code-do-developers-really-let-agents-write/
https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/
https://survey.stackoverflow.co/2025/ai
https://www.veracode.com/blog/2026-genai-code-security-report-ai-risk/
