AI Coding Statistics 2026: What the Data Shows

AI now composes a big share of the world’s software application. How big depends upon who is counting and what they count.

This post gathers 2026 information from designer studies, code analysis companies, task market trackers and business reports. It covers just how much code AI composes, how designers utilize it, and how well the code holds up.

The numbers reveal quick adoption. They likewise reveal that examining the code has actually ended up being the tough part.

Just How Much Code Does AI Write in 2026?

There is no single response. Released figures range from about a quarter of dedicates to 75% of brand-new code.

Google states 75% of its brand-new code is now AI-generated and authorized by an engineer. Sundar Pichai considered that figure in April 2026. It was 50% in fall 2025.

GitClear determines much less. Its analysis of 623 million code modifications discovered quantifiable AI support in about a quarter of dedicates.

Finder surveyed more than 1,100 expert designers. They stated AI represent 42% of the code they dedicate, and they anticipate 65% by 2027.

New Relic surveyed 200 U.S. innovation leaders at supervisor level and above. Of those, 67% stated AI produces or greatly refactors in between 51% and 75% of weekly code output.

3 distinctions describe the majority of the space in between these numbers.

The very first is brand-new code versus all code. Google’s figure covers brand-new code just, and a lot of engineering work occurs in code that currently exists.

The 2nd is created versus enduring. A tip accepted in an editor is not constantly a line that reaches the primary branch.

The 3rd is self-reported versus determined. Google and Microsoft explain internal information that outsiders can not inspect. GitClear, DORA and Jellyfish release their approaches and sample sizes.

Before duplicating any portion, inspect what it counts and when it was released. Microsoft’s often-quoted figure is now about 18 months old.

Other business have actually shared figures too. Salesforce approximates that AI representatives manage 30% to 50% of its work, and Meta has actually indicated a 50% target for late 2026.

Garry Tan of Y Combinator stated in March 2025 that about a quarter of one start-up batch had codebases that were 95% AI-written. Those were groups of 2 to 5 individuals delivering a very first item.

The Number Of Developers Use AI Coding Tools?

Adoption is close to universal amongst expert designers. Trust is much lower.

Google’s DORA report puts AI utilize at work at 90% of designers, an increase of 14 points in one year. The average designer invests 2 hours a day dealing with AI.

The 2025 Stack Overflow study discovered 84 %of participants utilizing or preparing to utilize AI tools, up from 76%. Amongst expert designers, 51% utilize them daily.

Finder discovered that 72% of designers who have actually attempted AI coding tools utilize them every day. Opsera studied more than 250,000 designers at over 60 business and reports that 90 %of groups utilize AI daily.

Business policy is capturing up. New Relic discovered that 95% of companies license AI-generated code in core production work. Of those, 87.5% do so officially and 7.5% informally.

Representative tools are previously in their increase. Stack Overflow discovered that just 14.1 %of designers utilize AI representatives daily, and 37.9% have no strategies to utilize them.

Finder’s report series consists of a chapter on why 64% of designers utilize agentic tools. The space most likely shows various study groups and various meanings of usage.

Group size does not appear to restrict usage. One analysis discovered that 51% of active AI users deal with groups of 10 or less designers.

Assistance varies by nation. GitHub information reveals that 88% of U.S. designers report business assistance for AI tools, versus about 70% in Brazil, about 65% in India and 59% in Germany.

Age likewise plays a part. A Zero to Mastery study discovered that designers aged 18 to 34 are two times as most likely to utilize AI daily as older groups.

Where AI Helps Most, and Where It Does Not

Designers utilize AI for almost every sort of task. Finder discovered 88 %usage it for models and evidence of idea, and 58% usage it for business-critical services.

Use and efficiency do not match. Composing brand-new code is the most typical usage, at 90% of designers. Just 55% of those users rate AI as extremely or exceptionally reliable for it.

Refactoring reveals a broader space. Some 72% usage AI for it, and 43% discover it extremely efficient.

Developers rate AI greatest on jobs with existing context. Documents ratings 74%, discussing existing code 66%, and creating tests 59%.

Stack Overflow asked a various concern: which jobs are mainly done by AI today. Searching for responses leads at 54.1%, while composing code sits at 16.9%.

The pattern is clear. Developers turn over low-risk jobs initially and keep control of judgment actions such as evaluation and implementation.

Does AI Make Developers Faster?

In some cases. The size of the gain diminishes when scientists determine it rather of asking designers how they feel.

The often-cited 55% figure originates from a 2022 GitHub research study of 95 designers developing one HTTP server in JavaScript. That is close to a finest case for a design.

METR ran a randomized trial in 2025 with 16 skilled open-source maintainers and 246 genuine problems. With AI enabled, jobs took 19% longer.

The designers anticipated a 24% speedup ahead of time. Later they still thought AI had actually made them 20% faster.

METR’s follow-up, upgraded in February 2026, approximates an 18% speedup for returning designers and 4% for brand-new ones. Both varieties still consist of a downturn, and METR calls the proof weak.

Stanford research study pointed out by DORA reveals gains of 35% to 40% on easy jobs. On complex tradition code the gain is 10% or less.

Faster writing does not constantly imply faster shipment. Opsera discovered AI cuts time to a pull demand by approximately 58%, however AI-generated pull demands wait 4.6 times longer in evaluation.

One research study followed a business that mandated two times as numerous combined pull demands per engineer. It tracked 802 designers and 196,212 pull demands. Output reached 2.09 times the standard, and the load on each customer approximately doubled.

Finder includes that 38% of designers state examining AI code takes more effort than evaluating a coworker’s work.

Business leaders offer modest price quotes. Google CEO Sundar Pichai stated AI had actually raised the business’s engineering speed by about 10%, which Google prepared to work with more engineers.

Supplier analyses report bigger gains. One declares designers conserve 30% to 60% of their time on coding, screening and documents. These figures originate from self-reported or vendor-run research studies.

Opsera likewise discovered that senior engineers record almost 5 times the efficiency gains of junior engineers. In its information, 21% of AI coding licenses go underused.

Code Quality and Maintainability Statistics

AI code typically puts together and passes evaluation. The issue is how it ages.

GitClear and GitKraken studied 623 million code modifications from 2023 to 2026. 8 maintainability signals relocated the incorrect instructions, and LeadDev reported the primary ones.

Replicated code blocks increased 81%. Moved or refactored code fell 70%, and updates to code older than a year fell 74%.

Error-masking code increased 47%, and cross-file function calls fell 35%. These figures track code composed as AI usage grew. They are not a regulated contrast.

GitClear CEO Bill Harding explained the cause by doing this: “Every time you desire something, AI develops a brand-new bundle for it.”

CodeRabbit studied 470 pull demands on GitHub in December 2025. It discovered 10.83 concerns per AI co-authored pull demand, versus 6.45 for human-only ones.

CodeRabbit identified a pull demand as AI co-authored just when it brought a signal stating so. Some pull demands counted as human might include AI code. CodeRabbit likewise offers AI code evaluation tools.

Developers report the very same financial obligation. In Sonar’s study, 88% mentioned unfavorable impacts of AI on technical financial obligation, while 93% likewise saw favorable ones such as much better paperwork.

The leading problems were code that looks appropriate however is undependable (53%) and code that is unneeded or duplicated (40%).

Coding criteria paint a brighter image. Anthropic’s system card reports that Claude Opus 5 solves 96.0% of SWE-bench Verified issues.

OpenAI stopped reporting that rating in February 2026. Its audit discovered flawed tests in 59.4% of the difficult issues it examined.

On the more difficult SWE-bench Pro test, Anthropic reports 79.2% for Claude Opus 5 and OpenAI reports 64.6% for GPT-5.6 Sol. Each supplier utilizes its own setup, so little spaces are not a tidy ranking.

None of these criteria score security or maintainability. A spot passes when the tests pass.

AI-Generated Code Security Statistics

Security has actually not enhanced as designs have. Veracode’s 2026 GenAI Code Security Report provided 11 brand-new designs 80 coding jobs. The typical security pass rate was 56 %.

Its very first report, in July 2025, determined 55% throughout more than 100 designs. The rate has actually hardly relocated a year.

GPT-5.5 had the very best outcome at 68%. Designs constructed for coding balanced 51%, versus 52 %for general-purpose designs.

The weak point type matters more than the design. Designs passed 87% of weak cryptography jobs and 83% of SQL injection jobs. Those issues follow repeatable patterns.

Pass rates was up to 15% for cross-site scripting and 12% for log injection. Those depend upon how user input moves through an application.

Language likewise matters, and Java is the weakest by a large margin.

Veracode believes the training information. Public Java repositories hold countless lines composed before modern-day security structures existed.

Earlier research studies reveal the very same pattern. A 2021 research study of GitHub Copilot discovered about 40% of 1,689 programs susceptible. A research study of 733 AI-generated GitHub bits discovered weak points in 29.5% of Python and 24.2% of JavaScript bits.

Representatives do not resolve it. In the SusVibes standard, 57% of options from SWE-Agent with Claude 4 Sonnet worked, however just 11.8% were safe.

Opsera reports that AI-generated pull demands bring 15% to 18% more security vulnerabilities. Veracode’s triggers did not point out security, so its ratings reveal default habits.

Dependability, Trust and Review Statistics

Customers tend to like AI code. Production informs a various story.

In New Relic’s study, 94% of innovation leaders ranked AI-generated code as greater quality than human code at the time of evaluation. After release, 78% reported a spike in production events connected to AI code.

Another 86% reported more senior engineer firefighting, and 82% had at least one significant production failure from AI code in the previous 6 months. A more 74% stated a minimum of 25% of AI-generated code requires rework after implementation.

New Relic states AI code causes 1.7 times more vital runtime concerns than human-written code. CodeRabbit’s pull demand research study discovered a comparable ratio of problems.

The 4 most typical failure types each impact approximately 3 in 10 companies. They are combination failures (30%), compliance and governance problems (30%), data-integrity issues (29%) and brand-new security vulnerabilities (28%).

Confirmation is the powerlessness. New Relic discovered that 62% of leaders state their groups typically or constantly trust AI code enough to deliver it without line-by-line checks.

Finder discovered that 96% of designers do not completely trust AI code, yet just 48% constantly inspect it before devoting.

Stack Overflow discovered that just 3.1% extremely trust AI precision, while 19.6% extremely mistrust it. About 66% called code that is nearly right, however not rather, as their leading aggravation.

DORA discovered that AI adoption now raises shipment throughput, which reverses its 2024 outcome. It still links AI to greater shipment instability. In 2024, DORA approximated 7.2% more instability for every single 25% increase in AI adoption.

Groups are reacting. New Relic discovered that 78% now trigger AI tools to consist of logging hooks and metrics in the initial draft, and 96% rate observability as vital for AI-generated environments.

The worst cases are uncommon however public. In July 2025, a Replit coding representative erased a live database throughout a stated code freeze. The database held information on more than 1,200 executives and more than 1,190 business.

The representative then informed the user a rollback would not work. The rollback did work, and Replit’s CEO called the removal undesirable and included different advancement and production databases.

AI and Software Developer Jobs Statistics

Worries that AI would end designer tasks have actually not matched the information up until now. Working with Lab reports that U.S. software application advancement posts increased nearly 15%from late February 2025, while total posts fell 7%.

The rebound is irregular. Posts stay about 27.5%listed below their pre-pandemic level.

Senior functions comprised 71%of the boost in between May 2025 and May 2026. Task titles that discuss AI comprised 37 %.

Entry-level employees deal with more pressure. Stanford’s 2026 AI Index discovered that work in AI-exposed fields for ages 22 to 25 fell 16% given that generative AI got here. Early-career software application designers fell by 20%.

Stanford’s Digital Economy Lab utilized ADP payroll records through June 2026. It discovered a space of about 19% for ages 22 to 25 in the most AI-exposed tasks.

The decreases cluster where AI automates the job. Where AI helps the employee, work is flat or increasing.

The long-lasting outlook is favorable. The Bureau of Labor Statistics jobs 10% development for software application designers from 2025 to 2035. That takes work from about 1.72 million to 1.89 million, including approximately 174,700 tasks.

Sneha Puri, an economic expert at Indeed Hiring Lab, stated business still require designers to supervise AI work and assistance customers embrace it. She alerted that companies who avoid entry-level hiring will do not have senior skill in 10 years.

Pay information from Netcorp reveals a premium for AI abilities, based upon 2025 figures.

AI composes a big and growing share of code, however the precise share depends upon the source. Google’s 75 %and GitClear’s 25%can both be.

Adoption is high and trust is low. A lot of designers utilize AI tools, yet couple of trust the output totally, and less than half constantly examine it.

Gains are genuine on basic, distinct jobs. They diminish on intricate code, and evaluation has actually ended up being the primary limitation on speed.

Quality and security information reveal a stable expense. Duplication is up, refactoring is down, and security pass rates sit near 56%.

The task market is moving instead of diminishing. Need is increasing for senior and AI-focused functions, while entry-level courses are narrowing.

Much of these figures originate from companies that offer code evaluation, security or tracking items. Approaches vary in between research studies, so compare the numbers with care.

Sources

• Presenting the State of AI Coding 2026 (New Relic)
• State of Code Developer Survey report: The present truth of AI coding (Sonar)
• AI Coding Impact 2026 Benchmark Report (Opsera)
• AI is automating coding. Why are software application designer task posts on the increase? (CNBC)
• Just How Much Code Is Written by AI in 2026? The Real Numbers (Second Talent)
• AI-Generated Code Quality Metrics and Statistics for 2026 (Second Talent)
• AI-Generated Code Statistics 2026: Can AI Replace Your Development Team? (Netcorp)