The interesting number this week is not a benchmark. It is a survey. JetBrains’ 2026 Developer Ecosystem sample, more than 15,000 developers in April and May, as reported by i-programmer on September 3, has Claude Code at 39 percent workplace use, up from 3 percent a year earlier. GitHub Copilot is at 21 percent, down from 29 percent. Awareness is tied at 79 percent. People know both. They are not opening both.
That would be a chart. Then Anthropic shipped background computer use for Claude Code and Cowork on Mac on September 2, Pro and Max only. Copilot still lives in the editor. Claude Code is trying to live on the desktop while you do something else.
This is not a rerun of Cursor vs Claude Code or Copilot vs Cursor. Cursor is the multi-model IDE in the middle. This comparison is the two names with the most name recognition, after one of them took the lead in a big survey and the other one did not ship a Mac agent this week.
What the JetBrains numbers actually say
90 percent of professional developers in that survey use AI coding agents at work at least weekly. 68 percent use them daily. The category is not experimental. The brand mix is.
Mikhail Bogdanov, writing on the JetBrains blog as quoted by i-programmer, called Claude Code the most widely adopted AI coding tool at work, used twice as often as Copilot. In the United States the Claude Code figure is 47 percent. Codex went from 0 to 16 percent, which puts it ahead of Cursor’s 12 percent use even though Cursor is better known (75 percent awareness vs Codex at 65, up from 11).
Surveys are not revenue. JetBrains respondents skew toward people who already care about tooling. Copilot still has the Microsoft 365 and GitHub Enterprise hose. A 21 percent use number is not a funeral. It is a loss of default status. For a long time Copilot was the thing your company already paid for. Claude Code is the thing people installed anyway.
If your procurement story is “we standardized on Copilot in 2024,” this survey is the slide a staff engineer will paste into Slack. You can still standardize. You should not pretend the floor is still 2024.
Background computer use is the product split
Copilot’s unit of work is a file, a chat, a pull request. Claude Code’s new unit of work, as of this week, is your Mac while you are not looking.
Zac Hall at 9to5Mac is right about the old failure mode. Computer use that steals the cursor is a demo. You cannot answer Slack. Background use is the difference between a babysitter and a contractor. OpenAI’s Codex and ChatGPT already did a version of this on the Mac earlier in the year. Anthropic is late to background, not late to computer use.
Android Authority’s writeup of Anthropic’s docs is the useful ladder. Claude tries connectors first. Then the in-app browser or Chrome. Only then does it ask to drive the screen and keyboard, once per session, and it waits if you are in the middle of something. Enable path: Settings, General, Computer use, in the Mac app. Windows is not invited. Linux is not invited. Free users are not invited.
That platform bet matters more than the blog-post tone. If your developers are on ThinkPads, this feature is a screenshot. If they are on MacBooks and already on Pro or Max, it is a reason to let Claude Code open Xcode, the browser, and the internal admin UI without standing over it.
Explainx’s roundup put the week’s three announcements next to each other: Anthropic on Mac GUI, OpenAI Codex on Windows and mobile control, Cursor on cloud agents that run on infrastructure you manage. Nobody is arguing whether an agent can click anymore. They are arguing where it clicks, and whether a human has to watch.
Copilot can still win the “I stay in VS Code / Visual Studio / neovim” job. It has not shown up to the “go click through Jira and the staging console” job on the same terms this week.
Limits, price, and the September 14 haircut
Claude Code is not getting cheaper. Pondero’s September 1 brief says Anthropic cuts the weekly limit 17 percent on September 14. If your team already bumps into weekly caps on agent loops, background computer use will not feel like a gift. It will feel like a faster way to burn the same bucket.
Copilot’s pricing is the opposite personality: bundled, seat-based, boring, and already in the GitHub invoice. You do not get a Mac-driving agent. You also do not get a surprise 17 percent smaller weekly allowance on a product that became the default because it felt unlimited enough.
Cursor sits in the Pondero note at $20 a month Pro, still the strongest multi-model editor in their view, with the OpenAI exit closing one of four model lanes rather than the core. That is context, not a third product in this table. If you need mixed models in one IDE, you are not choosing Copilot vs Claude Code. You are back in the Cursor comparison.
For this pair: Copilot is the seat you can explain to finance. Claude Code is the seat people already used, plus a Mac agent, minus headroom on September 14.
Editor gravity vs terminal gravity
Copilot is an extension with a chat panel and an agent mode that still assumes an editor. Claude Code is a terminal-first agent that grew a desktop app. VS Code vs Zed is the editor religious war. This is the interface religious war. Do you want the model in the gutter, or do you want the model in a session that can leave the gutter.
Teams that live in GitHub PRs, Actions, and Codespaces will keep Copilot because the comments appear where the review happens. Teams that live in local stacks, staging UIs, and “open the CSV in Numbers and fix the chart” will try background Claude Code because Copilot cannot see Numbers.
Neither is a compiler. Both will delete a file if you let them. Background permission once per session is better than silent cursor theft. It is not a substitute for a dirty git status before you walk away.
A practical split that matches the survey: keep Copilot on the org GitHub plan for inline complete and PR summary. Give Claude Code to the people who already expense Anthropic and who are on Mac. Measure weekly cap hits after September 14 before you make Claude Code the only agent. The JetBrains chart is a popularity contest. Your invoice is not.
Who should win on your machine
Pick Claude Code if you are on macOS, you already pay Pro or Max, you want an agent that can use the GUI without parking your mouse, and you can live with a tighter weekly limit in two weeks. The survey says a lot of your peers already voted that way at work.
Pick Copilot if your company is standardized on GitHub, your fleet is mixed OS, you need the thing to work in Visual Studio, or you refuse to run a desktop agent with screen control. The 21 percent use number is a decline, not an absence. Copilot is still the path of least procurement.
Pick neither as a monopoly if your threat model includes an agent that can click. Background use makes Claude more useful and more able to do something you did not watch. Copilot’s smaller surface is a feature for locked-down laptops.
A week-one test that is not a demo
If you are going to switch a team, do not start with “rewrite the auth service.” Start with a two-hour task that needs the GUI.
For Claude Code on a Mac: enable Computer use, give it a staging URL and a checklist, walk away, and see whether it asked for a screen takeover or finished in the browser. Time the weekly-limit warning. If you hit the cap on a toy task, September 14 will be worse, not better. Note whether it waited when you grabbed the mouse. Android Authority says it will wait. Your build of the beta might not.
For Copilot: same checklist, entirely inside VS Code, with Copilot Chat / agent mode against the repo only. No staging console. That is the honest comparison. If the task cannot be done without clicking a non-editor app, Copilot will lose by construction. If the task is “write the migration and open the PR,” Copilot still has home-field advantage because the comment lands on github.com where the review already happens.
Do not run both agents on the same dirty working tree. You will not know which one deleted the fixture. Commit or stash first. If you cannot explain the diff in a sentence, you are not ready to let either agent touch it.
Security review is part of the test, not a later phase. Background computer use is screen control with a permission prompt. Copilot’s blast radius is the repo plus whatever GitHub App scopes you already granted. A locked-down laptop with no admin and a no-computer-use policy is a Copilot machine. A founder MacBook that already has 1Password, Chrome, and staging bookmarks is a Claude Code machine. Pretending those are the same endpoint is how you get a bad rollout.
Codex at 16 percent in the same JetBrains survey is the spoiler. If your company is already paying OpenAI, you may get Windows and mobile computer use from that stack before you get it from Copilot. That does not make Copilot the winner. It makes “wait for Microsoft” a real option instead of cope. Watch the next GitHub Universe cycle. Do not freeze hiring of a Claude Code seat for six months on that hope unless you are actually on Windows.
Winner, for a Mac-heavy team that is already in the Anthropic ecosystem: Claude Code. The survey and the September 2 feature point the same direction. Copilot remains the safer default for mixed-OS enterprises that buy software by the seat and do not want a second agent with a cursor. If that is your shop, ignore the 39 percent and wait for Microsoft to ship a background computer-use story you can turn off in Intune.