# 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.

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.

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](/guides/testing/testing-ai-generated-code) 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)*.
- **Benchmarks:** vendor leaderboard claims (SWE-bench Verified, Terminal-Bench, Arena) are a separate race with their own harness, version and staleness caveats; [LLM Benchmarks Explained](/guides/evaluation/llm-benchmarks-explained-2026) says which number means what.
- **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](/guides/workflow/ai-code-review-workflow) that separate speed from [slop](/glossary/ai-slop).

## Continue exploring

- [Augment Code](/tools/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](/tools/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.

---

_Source: https://agentscamp.com/guides/concepts/ai-coding-statistics-2026 — Guide on AgentsCamp._
