AI Coding Assistant Statistics (2026): Adoption, Productivity, Security, and Market Data

You can watch an AI coding assistant turn a ticket into a plausible pull request in minutes, then spend the afternoon checking a dependency, reproducing an edge case, and untangling an almost-right implementation. The numbers for 2026 reflect that mixed experience: use is widespread, but measured productivity, trust, and security do not point in a single direction.The short answer: AI coding assistant statistics show high adoption, rapid shifts toward agentic tools, and unresolved evidence on real-world productivity. This report separates surveys, controlled experiments, direct security benchmarks, and third-party forecasts so you can cite a figure without overstating what it measures.

Key AI coding assistant statistics

• 90% of professional developers used AI coding agents at work at least weekly in May–July 2026, and 68% used them daily. This is a large-scale, globally representative JetBrains survey of more than 15,000 professional developers worldwide with regional quotas.

• 90% of DORA respondents reported AI adoption in 2025. Google’s DORA report surveyed about 5,000 technology professionals globally, including product managers as well as developers; it is a broader population than JetBrains’ developer-only sample.

• 84% of Stack Overflow’s 2025 respondents were using or planning to use AI in their development process. The primary Stack Overflow survey also reports that 51% of professional developers used AI tools daily. This is a survey measure, not product telemetry.

• Claude Code had about 39% workplace adoption in May–July 2026, versus 21% for GitHub Copilot. The figures come from the same 15,000-plus-developer JetBrains survey and represent reported tool adoption at work, not revenue share.

• Codex reached 16% workplace adoption in May–July 2026, up from 3% in January 2026. That is a change within JetBrains’ survey series, not a census of all developers.

• In GitHub’s 2022 experiment, 95 professional developers completed an artificial JavaScript task 55% faster with Copilot. Average completion time was 1 hour 11 minutes with Copilot versus 2 hours 41 minutes without it, according to GitHub’s controlled experiment.

• In METR’s early-2025 real-task experiment, 16 experienced open-source developers were 19% slower when using AI tools. The estimated slowdown’s 95% confidence interval was 2% to 39%, as reported by METR.

• More than 80% of DORA respondents said AI enhanced their productivity in 2025. This is self-reported perception among approximately 5,000 technology professionals, not timed task completion data.

• AI-generated code averaged a 56% security pass rate in Veracode’s four-year benchmark. The Veracode report tested more than 100 models on 80 security tasks per model.

• 69% of organizations in Aikido’s 2026 survey said they had found vulnerabilities introduced by AI-generated code. The survey covered 450 security leaders, developers, and AppSec engineers in Europe and the United States, according to Aikido Security.

• Only 29% of Stack Overflow’s 2025 respondents trusted AI output to be accurate. Stack Overflow’s survey summary says 66% spent more time fixing AI-generated code that was “almost right.”

• Third-party forecasts range from a $4.1 billion AI code-assistant market in 2026 to a $14.18 billion enterprise AI coding-agent market in 2026. Those are different forecast products with different scopes, so they are not competing measurements of one market.

Get the Mobile Testing Playbook Used by 800+ QA Teams

Discover 50+ battle-tested strategies to catch critical bugs before production and ship 5-star apps faster.

Email addressGet Instant Access

100% Free. No spam. Unsubscribe anytime.

AI coding assistant adoption

The best current workplace adoption figure is JetBrains’ 2026 result: 90% of its more than 15,000 surveyed professional developers used AI coding agents at work at least weekly, and 68% used them daily during May–July 2026. The survey is a strong indicator of professional-developer use because it is large and globally structured, but it remains self-reported survey evidence.DORA reached the same 90% headline percentage through a different question and population. In its 2025 survey of roughly 5,000 technology professionals, DORA reported 90% AI adoption and a median of two hours per day working with AI. DORA explicitly includes people “from developers to product managers,” so it should not be presented as corroboration of JetBrains’ developer-only weekly-agent-use statistic.Stack Overflow provides a third lens. Its 2025 survey says 84% of respondents were already using or planning to use AI tools in the development process, while 51% of professional developers reported daily use. The first statistic includes future intent; if you need a current-use measure, do not cite it as though all 84% were active users.The useful conclusion is not that there is one definitive adoption rate. It is that several large surveys show AI is embedded in many development workflows, while measuring different populations, usage thresholds, and periods.

AI coding assistant tool usage in 2026

JetBrains supplies the clearest available large-scale comparison of individual AI coding assistants. Its figures describe reported adoption at work among more than 15,000 professional developers worldwide in May–July 2026. They are not active-user counts or global market share.

Claude Code’s approximately 39% adoption was up from 18% in JetBrains’ January 2026 reading, and 31% of respondents named it as their single most-used AI coding tool. Copilot fell from 29% a year earlier to 21% in the May–July fieldwork.Codex moved from 3% adoption in January to 16% in May–July 2026, while Cursor moved from 18% to 12% over the same period. Regional results also differ: Claude Code reached 47% adoption in the United States, and Google Antigravity reached 15% in India. A global average is useful for orientation, but it can conceal local tool preferences.Tool adoption is moving faster than most annual surveys can capture. If you are selecting tools, pair survey rankings with a workflow-specific evaluation: what the assistant can access, where it is allowed to act, and what review controls apply. For the adjacent risk profile of prompt-led development, see Quash’s research on vibe coding adoption and verification risks.

Productivity evidence points in different directions

The major AI coding productivity studies do not support one universal percentage. They test different tools, task types, developer populations, and outcomes.

GitHub measured a 55% gain on an artificial task

GitHub’s 2022 randomized experiment assigned 95 professional developers to write an HTTP server in JavaScript. The Copilot group completed the task in an average of 1 hour 11 minutes, compared with 2 hours 41 minutes for the control group; GitHub reported a 55% faster completion time, with a 95% confidence interval of 21% to 89%.This is controlled experimental evidence, not a broad survey. It tested autocomplete-era Copilot on a defined exercise, so it cannot establish the impact of 2026 agentic assistants on a production repository, code review, or maintenance work.

DORA recorded perceived gains, not timed output

In DORA’s 2025 survey, more than 80% of respondents said AI enhanced their productivity and 59% reported a positive influence on code quality. These figures describe what surveyed technology professionals reported feeling and observing in their work.Self-reported productivity is valuable evidence about user experience and adoption. It is not interchangeable with a timed experiment, because respondents may judge speed from drafting, ideation, or task completion without measuring downstream debugging and review time.

METR measured a 19% slowdown on developers’ own tasks

METR ran a randomized controlled trial with 16 experienced open-source developers working on their own repositories in February–June 2025. It found that using early-2025 AI tools made the participants take 19% longer, with a 95% confidence interval from 2% to 39% slower.That finding does not prove that AI slows every developer. It is a small, real-task experiment involving experienced contributors in familiar codebases—conditions that differ substantially from GitHub’s one-off JavaScript exercise.

A follow-up exposed the difficulty of creating an AI-free control group

In February 2026, METR changed its follow-up design after developers increasingly declined to do portions of their work without AI. Between 30% and 50% of developers chose not to submit some tasks without it. METR called its new estimates unreliable because the participation pattern created selection effects; the confidence intervals for its preliminary estimates included zero.That design problem matters beyond the result itself. It shows how difficult it is becoming to run a clean AI-versus-no-AI comparison in normal developer workflows. It still does not yield a reliable agentic-era estimate of real-task productivity.

Code quality, security, and trust statistics

Adoption and reported speed do not remove the need for review. The available evidence includes both direct testing of generated code and surveys of people who ship it; those evidence types answer different questions.Veracode’s 2026 benchmark tested more than 100 AI models over four years using 80 security tasks per model. The average security pass rate was 56%, with no improvement in the average over the period. In its Summer 2026 dataset, the top model reached a 68% pass rate; the report also found code-focused models averaged 51%, versus 52% for general-purpose models. This is an empirical benchmark of tested outputs, not a survey of production repositories.Aikido’s 2026 survey reports organizational experience instead. Of 450 surveyed security leaders, developers, and AppSec engineers in Europe and the United States, 69% said their organization had found vulnerabilities introduced by AI-generated code, and one in five reported a serious incident linked to it. Those are self-reported events, not independently verified incident counts.Safeguard’s April–May 2026 survey of 1,412 developers, AppSec engineers, and engineering leaders in North America and Europe found that 91% used an AI coding assistant at least weekly and 58% used one every working day. Yet only 34% trusted AI-generated code to be secure by default, while 63% reported shipping AI-suggested code without fully reviewing it. The same Safeguard report says only 26% of respondents’ organizations had a formal policy requiring security review specifically for AI-generated pull requests.Stack Overflow’s 2025 results point to the same tension: its 49,000-plus global respondents showed 29% trust in AI accuracy, down from 40% in previous years. Trust is perception data, while Veracode’s pass rate is a task benchmark, but together they support a practical rule: treat generated code as code that still needs normal review, testing, and dependency checks. That is especially relevant when AI-generated change volume reaches QA; Quash’s AI testing statistics and adoption data covers the testing-side implications.

AI coding assistant market forecasts

No official statistical agency publishes a single AI coding assistant market total. The available numbers are third-party forecasts, and each firm draws the category boundary differently.

Alora Advisory’s March 2026 forecast defines its broad category to include IDE autocomplete, in-IDE chat, agentic execution engines, end-to-end application generators, and enterprise developer-productivity platforms. It estimates approximately $6 billion in 2024 and projects approximately $36 billion by 2030, a 34–36% CAGR.Mordor Intelligence forecasts a narrower enterprise AI coding-agent category: $10.42 billion in 2025, $14.18 billion in 2026, and $45.83 billion by 2031, with a projected 26.44% CAGR from 2026 to 2031. Its estimate includes a different product and buyer scope from Alora’s broad software-engineering market.Future Market Insights uses the narrowest label here, forecasting the AI code-assistant market at approximately $4.1 billion in 2026 and approximately $6.9 billion in 2036, at a projected 5.3% CAGR. The range is not a contradiction to average away: it is the result of different market definitions. Cite the forecaster, scope, base year, and forecast horizon whenever you use these figures.

What the public statistics still cannot tell you

There is no reliable agentic-era real-task productivity benchmark

The best-known positive experiment studied autocomplete on an artificial 2022 task. METR’s real-task result studied early-2025 tools and a small group of experienced open-source developers. DORA measures self-reported benefit. Public evidence still lacks a large, reliable study of how 2026 agents such as Claude Code and Codex affect real work across task types.

There is no shared market definition

Assistants, coding agents, application generators, and broader developer-productivity platforms are often placed under the same market label. The forecast table shows why a single uncaveated market-size number is not a usable benchmark.

We do not know which tasks gain time and which add review work

Current studies do not provide a consistent task-level breakdown across boilerplate generation, debugging, architecture decisions, code review, and long-term maintenance. That is the missing evidence you need for staffing and workflow decisions.

Long-term maintainability remains unmeasured

A security pass rate captures performance on test tasks, not the maintenance cost of years of AI-assisted changes in a living codebase. Public research has not yet established whether faster generation changes review burden or defect patterns over the long term.No first-party Quash telemetry, recurring bug pattern, customer language, or original experiment covers AI coding assistant adoption, productivity, market size, tool usage, or code quality. What would materially improve this record is a reproducible, agentic-era study that measures task completion, review time, defects, and maintainability on real repositories rather than treating a single speed figure as the whole outcome.

Methodology

This report prioritizes first-party survey publishers and direct experiments. Survey figures are labeled as survey responses; they do not represent all developers or all product users. JetBrains surveyed more than 15,000 professional developers worldwide in May–July 2026, DORA surveyed approximately 5,000 technology professionals in 2025, and Stack Overflow surveyed more than 49,000 respondents in 2025.The productivity studies are deliberately kept separate. GitHub used an artificial JavaScript task with 95 professional developers, while METR used real tasks with 16 experienced open-source contributors. Their results should not be averaged.Security evidence also has separate meanings. Veracode measures outcomes on defined security tasks; Aikido and Safeguard report what surveyed security and engineering practitioners said occurred in their organizations. Market figures are third-party projections, not company-disclosed category revenue.

Conclusion

AI coding assistant statistics in 2026 support a narrower, more useful conclusion than “AI makes developers faster.” Developers report widespread use, and JetBrains’ survey shows that tool preferences are shifting quickly toward agentic products. But the productivity record spans a controlled 55% gain, strong self-reported benefits, and a controlled 19% slowdown on real work.Use the evidence that matches your decision. For adoption, cite the survey and its population. For productivity, name the task and study design. For security, distinguish benchmark results from reported incidents. For market sizing, name the forecaster’s scope. Your next decision is not whether to allow AI coding assistants—it is where to apply them, and what review and testing evidence you require before their output reaches production.