Skip to content
Level up your prompts with SurePrompts — curated prompts for every workflow.
agentscamp

AI Coding Statistics 2026: The Numbers That Are Actually Sourced

How much code AI writes, who uses the tools, and what it does to quality — every statistic dated and traced to its primary source, updated on a cadence.

By Imtiaz RayhanUpdated 4 min read

On this page
  1. How much code does AI write?
  2. Who's using the tools
  3. The tool race, by sourced metric
  4. What it does to productivity and quality
  5. Continue exploring

The sourced numbers, September 2026: Google says 75% of new code is AI-generated; 84% of developers use or plan AI tools (Stack Overflow); GitHub Copilot hit 50M users; Claude Code passed $2.5B run-rate; SpaceX closed its $60B all-stock Cursor acquisition on August 14; and the METR RCT found experienced devs 19% slower. Every figure dated and sourced.

Key takeaways

  • The code-share headline: Google reports 75% of its new code is AI-generated and engineer-approved (April 2026, up from ~25% in late 2024); Microsoft reported 20–30% a year earlier.
  • Adoption is near-saturation: 84% of developers use or plan AI tools (Stack Overflow, 49k respondents), 90% of tech professionals use AI at work (DORA), 85% regularly (JetBrains) — the question moved from whether to how.
  • Agents crossed the chasm in 2025–26: 31% of developers used them (SO 2025), 55% of engineers regularly (Pragmatic Engineer 2026, senior-skewed sample), with Claude Code the most-used and most-loved tool in that survey — and Microsoft says 1 in 3 GitHub pull requests now involves an agent (July 2026).
  • The productivity evidence is genuinely mixed: DORA 2025 found AI adoption finally correlating with delivery throughput (but still hurting stability); the METR RCT found experienced devs 19% SLOWER while believing they were faster.
  • Trust lags usage: 46% of developers distrust AI output accuracy (up from 31% in 2024); the top frustration is AI answers that are 'almost right, but not quite' (66%), with debugging AI-generated code close behind (45%).
  • The category's biggest deal yet is done: SpaceX completed its all-stock acquisition of Cursor (Anysphere) on August 14, 2026 at a $60B implied equity value, per its SEC 8-K — two months after the June 16 announcement.

AI-coding statistics are mostly laundered guesses — numbers that trace to an SEO listicle citing another listicle. This page is the opposite: every figure below is dated, sourced, and labeled (primary / survey / reported), verified September 1, 2026, and refreshed on a cadence. Numbers we couldn't trace are omitted.

How much code does AI write?

  • 75% of new code at Google is AI-generated and engineer-approved — Sundar Pichai, Cloud Next, April 2026 (primary). The trajectory: >25% (Oct 2024) → "more than 30%" (Apr 2025) → ~50% (fall 2025) → 75%.
  • 20–30% of code in Microsoft's repos "written by software" — Satya Nadella, April 2025 (reported).
  • ~4% of all public GitHub commits authored by Claude Code, per an external analysis cited in Anthropic's Series G announcement, February 2026 (primary, second-hand analysis).
  • Context for scale: GitHub logged nearly 1 billion commits in 2025 (+25% YoY), with 1.1M+ public repos importing an LLM SDK (+178% YoY) — Octoverse, October 2025 (primary).

Who's using the tools

  • 84% of developers use or plan to use AI tools (76% in 2024); 51% of professional developers use them daily — Stack Overflow Developer Survey, 49,000+ respondents, July 2025 (survey). (The 2026 survey opened June 23; results were not yet published at our September 1 check — "2026 survey" numbers circulating online are 2025 figures relabeled.)
  • 90% of tech professionals use AI at work (+14 pts YoY), median 2 hours/day with AI — DORA, ~5,000 surveyed, September 2025 (survey).
  • 85% regularly use AI tools; 68% expect AI proficiency to become a job requirement — JetBrains State of the Developer Ecosystem, 24,534 devs, October 2025 (survey).
  • Agents specifically: 31% of developers used AI agents in 2025 (SO); by early 2026, 55% of engineers used agents regularly — 63.5% among staff+ — Pragmatic Engineer survey, 906 respondents, (survey; self-selected, senior-skewed sample).
  • Trust lags: 46% distrust AI output accuracy (31% in 2024); the top frustration is AI output that's "almost right, but not quite" (66%), with time spent debugging AI-generated code close behind (45%) — SO 2025. The verification stack exists for a reason.

The tool race, by sourced metric

  • Preference: Claude Code ranked most-used and most-loved (46% most-loved, vs Cursor 19%, Copilot 9%) — Pragmatic Engineer, March 2026 (survey).
  • Scale: GitHub Copilot reached 50M users (Microsoft FY26 Q4 earnings, July 29, 2026, primary), up from 20M a year earlier; the last disclosed paid figure is 4.7M subscribers, +75% YoY (January 2026) — the two are not like-for-like. The same call put GitHub at 225M users with 1 in 3 pull requests involving an agent; ~80% of new GitHub users adopt Copilot in week one (Octoverse, primary). Google's Antigravity passed 2.4M weekly active users (Alphabet Q2 call, July 22, 2026, primary).
  • Revenue: Claude Code hit $1B run-rate six months after GA (December 2025) and >$2.5B by February 2026, with enterprise over half of it — Anthropic (primary; no Claude Code breakout since). Cursor's annualized revenue climbed from $2B (February) to $3B (late April) — Bloomberg — to ~$4B (early June 2026) — Forbes (reported); SpaceX announced its acquisition of Cursor (Anysphere) on June 16 and closed it on August 14, 2026, issuing 389.3M Class A shares at a $60.0B implied equity value (primary, SEC 8-K). OpenAI's Codex passed 5M weekly active users (June 2, 2026, primary); its product lead posted 25M "active users" on August 31 (primary, exec social post — no time window stated, includes non-developer knowledge workers).
  • The builders: Lovable confirmed $500M ARR as of June 2026 (up from $400M in February) and raised a $400M Series C at a $13.3B valuation on August 12, 2026 (reported, company-confirmed / primary); Bolt went $0→$20M ARR in two months post-launch (reported, founder on record).

What it does to productivity and quality

The honest section. For: DORA 2025 found AI adoption positively associated with delivery throughput for the first time (a reversal from 2024), and >80% of practitioners perceive productivity gains (survey). Against: the METR randomized controlled trial — the only RCT on experienced developers and real tasks — measured them 19% slower with early-2025 tools, while they believed they were ~20% faster (July 2025, primary; a February 2026 follow-up on late-2025 tools found point estimates tilting toward a modest speedup, but with confidence intervals crossing zero — METR calls it only very weak evidence). Quality: GitClear's analyses — 211M changed lines through 2024, and a June 2026 follow-up over 623M changes — find code duplication rising sharply in the AI era (block duplication up 81% from 2023 to 2026 year-to-date; refactoring down to 3.8% of changes) (primary, vendor research — affiliation disclosed); DORA still finds AI adoption negatively associated with delivery stability.

The synthesis this page stands behind: adoption is real and enormous; measured productivity is conditional — on task, skill, and above all on the verification practices that separate speed from slop.

Continue exploring

  • Augment Code — AI coding assistant built for large, real-world codebases — a Context Engine that indexes the whole repo, with agents, chat, and completions in IDEs and a CLI.
  • Trae — Trae is an AI-native IDE from ByteDance — a VS Code-style editor with a built-in Builder agent and an autonomous SOLO mode that writes code across a project.
ShareShare on XShare on LinkedIn
View as Markdown

Frequently asked questions

What percentage of code is written by AI in 2026?
The best-sourced datapoint: Google's CEO stated in April 2026 that 75% of the company's new code is AI-generated and approved by engineers — up from 'more than 25%' in late 2024 and 'well over 30%' in early 2025. Microsoft reported 20–30% in April 2025. Industry-wide there's no single credible figure; one external analysis cited by Anthropic put Claude Code alone at ~4% of all public GitHub commits by February 2026.
Does AI actually make developers faster?
The honest answer is contested. Self-reports say yes (>80% in DORA 2025 perceive productivity gains; 69% of agent users in Stack Overflow's survey). The one randomized controlled trial — METR, July 2025 — found experienced open-source developers 19% slower with early-2025 tools on real tasks, while believing they were ~20% faster (METR's February 2026 follow-up on newer tools found point estimates tilting toward a modest speedup, but the confidence intervals cross zero and METR calls it only very weak evidence). DORA's org-level data found AI adoption associated with higher delivery throughput but lower stability. Perception, task type, and skill clearly mediate; treat blanket productivity claims skeptically.
What's the most popular AI coding tool in 2026?
By the Pragmatic Engineer survey (906 engineers, early 2026, senior-skewed): Claude Code is both most-used and most-loved (46% 'most loved' vs Cursor's 19% and Copilot's 9%). By raw scale, GitHub Copilot's 50M users (Microsoft, July 2026; its last disclosed paid count was 4.7M in January) remain the biggest footprint, and OpenAI's Codex reported 5M+ weekly active users in June 2026 — its product lead posted 25M 'active users' on August 31, a looser metric. Different metrics crown different tools — usage breadth, paid depth, and developer preference are three different races.
Where do these numbers come from?
Every statistic on this page carries its source, date, and a quality label — primary (the organization's own announcement or data), survey (named methodology), or reported (credible press citing a primary). Widely-circulated numbers we could not trace to a credible source are deliberately omitted.

Filed under

Coding & LanguagesAI Agents & SystemsFor AI Engineers

statistics · ai-coding · data · adoption · research

Related