Who uses which AI coding agent, counted five ways

JetBrains asks professional developers which AI tools they use at work, with quotas by region so one country cannot swamp the result. It asked in January and again from May to July 2026, the second time of more than 15,000 people. Grey is January, blue is mid-year.

Claude Code more than doubled, to 39% worldwide and 47% in the United States. Codex went from 3% to 16%, and the share who had heard of it from 27% to 65%. Copilot and Cursor are still widely known, Copilot by 79% of developers, but fewer use them at work. OpenCode reached 7% with no large company behind it.

The picture moves by country. In India, Google's Antigravity is used by 15% of developers, up from 10% in January and level with Cursor there, against 6% worldwide.

Stack Overflow's 2026 survey, a larger and self-selected sample, agrees on the top two: of developers who use coding agents, 66% named Claude Code and 59% GitHub Copilot.

OpenRouter sells access to hundreds of models through one key, and publishes how many tokens each app sends through it. It is the only public measure of what agents actually do, rather than what people say. In the 30 days to 13 October, the top thirty coding apps sent 192.3T tokens. The top 16:

This measure leans one way, and it is worth knowing which. An agent that lets you pick any model, like Hermes, Kilo Code, Cline or pi, is exactly what a router is for, so they top it. Cursor and Copilot sell their own plans with the models included, so their traffic rarely passes through here: Cursor is #16, and Copilot is absent. Claude Code's #2 counts only the people who run it on models bought through OpenRouter, not those on an Anthropic plan.

The change since January is the steepest number on this page. Claude Code sent 0.44T tokens in January and 32.6T in the last 30 days, 74× as much. Roo Code went the other way, from 0.86T and second place in January to 0.23T; its repository was archived in May.

Many agents sign their work: a co-author line on a commit, a bot account on a pull request. amplifying.ai counts those marks across public GitHub. On 12 October it attributed 188M commits and pull requests to 30 agents.

Claude Code holds 81% of all signed work. In a sample of the 1,999 most recent public pull requests, 32.5% carried an agent's mark and 29.8% carried Claude Code's alone.

Two blind spots. An agent used inside an editor usually commits under the developer's own name and leaves no mark, which undercounts Cursor and Copilot. And a squash merge strips the marks, so every figure here is a floor. Codex is counted in pull requests rather than commits, so its row is not like for like.

Codex's 52× is the largest move, and also the one to read most carefully: a download is not a user. Build servers reinstall a package on every run, and an agent that runs in CI is downloaded thousands of times a day by nobody in particular. Claude Code also ships a native installer that npm never sees, so its count is low. What npm is good for is direction. Gemini CLI, at 1.1× its January level, is flat while almost everything around it grew.

This page is the kind of document hedwigAI is for: a table built from public sources, every figure cited, that goes stale the week it is written unless something re-reads the sources. A hedwigAI workbook does the re-reading on a schedule you set, keeps the sources beside each number, and asks you before it changes what a column means.