According to JetBrains’ May to July 2026 survey of more than 15,000 developers, AI agents now write an average of 47% of professional developers’ code. Developers also estimate 38% is AI-assisted and 27% is written by hand. JetBrains calculates those figures from answer-bucket midpoints, which is why the three numbers add up past 100%.
Over half of developers now write less than 20% of their code by hand. One in five writes no code without AI. About 22% get more than 80% of their code from agents.
Tool share has shifted fast
Claude Code is now used at work by 39% of professional developers, up from 18% in January 2026. GitHub Copilot went the other direction, dropping from 29% to 21% over the same period. OpenAI Codex grew roughly fivefold in work usage, from 3% to 16%. Cursor fell from 18% to 12%.
| Tool | Work use, Jan 2026 | Work use, May–Jul 2026 |
|---|---|---|
| Claude Code | 18% | 39% |
| GitHub Copilot | 29% | 21% |
| OpenAI Codex | 3% | 16% |
| Cursor | 18% | 12% |
| Google Antigravity | 6% | 6% |
Claude Code is the primary tool for 31% of developers overall. In the US that figure reaches 47%.

Daily use and adoption
90% of professional developers used AI coding agents at least weekly in May to July 2026. 68% used them daily. Earlier in the year, JetBrains’ AI Pulse survey found 74% had adopted specialized developer tools, not just general chatbots.
Stack Overflow’s self-selected 2025 survey of 49,000+ respondents tells a different story: only 14.1% used AI agents daily, and 37.9% had no plans to adopt them. The two surveys sample different populations and ask different questions, so the gap does not represent a single measured trend.
Trust and security: the gaps that matter
96% of developers do not fully trust AI-generated code, per Sonar’s January 2026 survey of 1,100+ developers. Yet only 48% always verify AI code before committing. 38% say reviewing AI code takes more effort than reviewing a colleague’s work. 35% access AI coding tools through personal accounts rather than work-approved ones.
Veracode’s July 2026 report tested 100+ models and found a 56% average security pass rate. About 44% of coding tasks produced code with a known vulnerability when no security-specific prompting was used. GPT-5.5 led the tested set at 68%. Java averaged 30% and ranked last among languages tested.
⏱️ Is AI actually faster?
METR’s early-2025 randomized trial found that AI-allowed tasks took 19% longer than those done without AI. A late-2025 follow-up pointed toward speedups, but METR calls that data unreliable. Between 30% and 50% of participants said they withheld tasks they would not want to do without AI, which likely pulls the measured speedup down. A separate METR survey of 349 technical workers found a median self-reported 1.4x to 2x increase in the value of their work from AI tools, though METR flags reasons to question that size.
Revenue
Claude Code’s run-rate revenue passed $2.5 billion by February 12, 2026, more than double its level at the start of the year. Enterprise use made up over half of that revenue. Anthropic cited a SemiAnalysis estimate that Claude Code authored 4% of public GitHub commits.
GitHub Copilot had 4.7 million paid subscribers in January 2026, up 75% year over year. By July 29, 2026, Copilot counted 50 million users across GitHub’s 225 million total. Copilot revenue grew more than 60% quarter over quarter after usage-based pricing launched in June. Satya Nadella said 1 in 3 GitHub pull requests now involves an agent.
Google reported 75% of its new code as AI-generated and engineer-approved as of April 2026, up from 25% in October 2024 and 50% in fall 2025.

