Claude Code is the default coding agent at work
August 21, 2026
The default coding agent at work is now Claude Code, at least among the professional developers JetBrains counted from May to July 2026. 39% used it at work. 31% named it their single most-used AI coding tool. That conversion, about 80%, is the whole story.
Awareness is a tie. Copilot still sits at 79% mindshare, same number as Claude Code on the published chart. The license on a lot of laptops is still Copilot.
The chair people actually sit in is not.
Default here does not mean majority. 69% of that sample still live in some other tool as their main one. If your shop bought Copilot in 2024 and never revisited the SKU, the survey is describing a different building than the one you walk into.
How JetBrains counted a default#
Can you trust a JetBrains survey to name a workplace default? Only if you read the question they asked, and the pile they reweighted.
The May to July wave prints N=15,509 on the chart axis. Heise reported the same headcount. "Professional developers" means people who code in a listed job role, and about 90% of the sample is Developer, Programmer, or Software Engineer.
Eight languages. Regional quotas. Then a reweight for region, employment status, programming language, and familiarity with JetBrains products.
The work-usage question is not "what did procurement buy." It is which of these AI tools you use for coding and other development-related tasks at work, with multiple answers allowed, and only a subset of tools shown.
- Copilot sits on that same chart as Claude Code
- Cursor and Codex do too
- ChatGPT does not, which is a limit later
This part is boring on purpose. The drama is in who converted.
January 2026 is a different wave, the AI Pulse of more than 10,000, promoted without mentioning AI so the sample would not fill with fans. The 2025 methodology is the honesty document JetBrains still points at. Lingering bias is named.
JetBrains users may have been more willing to finish. Raw 2026 microdata is not out yet. Self-report. Not seat telemetry. Not git trailers.
If you need a sensor in the repo, this is not that post.
Conversion is the default, not majority#
Those bars are multi-select work usage. The 31% most-used number is a different, exclusive question, printed only for Claude Code. The bars cannot show conversion by themselves.
If only 39% use Claude Code at work, why call it the default? Because default, in this survey, is the tool people make their main one, not the tool they have heard of, and not the tool that shows up on an invoice.
JetBrains writes it flat. Claude Code became by far the most widely adopted AI coding tool at work. Used twice as often as Copilot.
39% worldwide in May to July, up from 18% in January. 47% in the United States. 31% most-used. Almost 80% conversion from regular work usage to the single most-used tool.
39 divided by 31 is not the math. 31 divided by 39 is. People who pick Claude Code tend to live in it.
The 90% weekly and 68% daily numbers are the category, not the product. Agents as a class are already at work. The fight is which chair.
Claude Code's own docs call it an agentic coding tool that reads the repo, edits files, and runs commands, in the terminal, the IDE, a desktop app, and the browser. That is a different object than autocomplete that hijacks Tab. The survey put both families on one chart.
The conversion number is how you tell them apart. You can still lose this bet if your team is the 69% whose most-used tool is something else. Plurality is not a mandate. It is a weather report.
Copilot kept the license and the awareness tie#
Did Copilot actually lose, or only lose the survey headline? Usage fell. Fame did not.
The purchase order often did not.
Work usage sat at 29% a year ago, 29% again in the January Pulse, then 21% in May to July. Awareness is still 79% worldwide, and 86 to 90% in Europe, the UK, and the US. 39% of Copilot users run it in JetBrains IDEs, which is a fun detail in a JetBrains paper and also a real install base.
January Pulse had Copilot at 40% inside companies with more than 5,000 people. That is the building with a vendor review cycle measured in quarters. HN asked, in March, whether Copilot was still relevant in the enterprise.
The top reply was the policy sentence. Copilot was the only AI allowed. Agent mode had already wrecked a working tree. Claude Code was "some talk."
You will ship a Copilot SKU and still watch the people who can choose live in a different CLI. That is not hypocrisy. That is two clocks.
Procurement is annual. Conversion is weekly.
Stack Overflow's 2025 survey is the year-ago texture, not a 2026 rebuttal. Agents were "not yet mainstream." ChatGPT and Copilot were the out-of-the-box leaders among people who used agents at all.
Then the first half of 2026 happened, and the JetBrains chart bent. Keep Copilot if it is cheap autocomplete and the Tab key is still a habit. Do not confuse that seat with the agent default.
They are not the same chair.
Cursor got more famous while people left#
If Cursor is in every feed, why is work usage falling? Because a poster is not a chair.
Awareness went from 69% in January to 75% in May to July. Work usage went from 18% to 12%. China is the ugly split inside the ugly split, 28% down to 16%.
JetBrains printed that drop in the same breath as the global decline, which is the honest way to do it. This is the column the reprints skip. Heise and daily.dev ran the Claude Code lead.
The Cursor line is the one that should make a lead nervous. You can buy the story that the tool is everywhere, look at the mentions, and still be losing the only number that is a default.
Fame without conversion is a tax. Design partners, conference clips, a model name on a keynote slide. Meanwhile reported work usage is falling.
The survey does not say why. It does say the direction of the arrow. That is enough to stop ranking tools by mention volume.
SpaceX closed the Cursor acquisition on August 15, per TechCrunch. That is after the May to July window. Treat it as a caption.
GPU fleet, $60 billion, joining SpaceX. None of that rewinds a usage drop that already printed. If the close changes the 2027 chart, that is a later post.
You'll hit this if your strategy meeting is still ranking tools by who is in the timeline. The timeline is awareness. The chair is usage, then most-used.
Cursor is gaining the mentions and losing the chair. A clean ranking post would give that a paragraph and move on. The split is the thing a competitor writeup dodges, because it complicates the Claude Code won headline they already copied.
Won on conversion. Lost nothing on Cursor's mindshare. Those can both be true in one quarter, and that is the mess you actually manage.
Codex is the chase, OpenCode is the leak#
Should a shop wait for Codex, or jump to OpenCode, instead of standardizing? Codex is the sprint. OpenCode is the side door.
Neither is the default yet.
Codex went from 3% work usage in January to 16% in May to July, about 5x, the fastest climb on the chart. Awareness jumped from 27% to 65%. That is a lot of people who now know the name, and a smaller group who report using it at work.
Conversion is still Claude Code's trick. OpenCode reached 7% work usage with 42% awareness and no big company name behind it. The OpenCode repo is the public artifact, a couple hundred thousand stars and still getting pushes.
Mindshare without a parent is the leak in a market that otherwise looks like a vendor race. A same-task bench already lived here. OpenCode vs Claude Code on four graded tasks both hit 4/4.
Paths diverged on tools, wall time, and reported tokens. Product lane, not a lab that proves share. Share is this survey. Pass rate is that post.
Do not mash them. Codex vs Claude Code on quota is also already measured, in the same unit across 28 days. That post will not tell you who is default at work.
This one will not tell you which weekly cap burns first. Read both if you are picking a seat.
Antigravity held 6% usage with awareness up to 47%, India at 15%. JetBrains AI in IDEs and/or Junie around 9%. The IDE vendor is hosting Claude Agent, Codex, Copilot, and OpenCode in the same chat, plus ACP for the rest.
That is not a ranking. That is an admission that the default is portable. Waiting for Codex because the slope is steep is a reasonable hedge. Treating a hedge as the current default is how you miss the 31% already sitting in Claude Code.
What 15,509 self-reports cannot say#
What would make this default claim fold? A different question, a different sample, or the missing card on the table.
The May to July charts are a subset. ChatGPT is not on them. The January Pulse still had the ChatGPT chatbot at 28% of developers for coding tasks at work.
JetBrains spent May asking, on X, whether anyone remembered ChatGPT as last year's most popular coding tool. This post cannot crown Claude Code over ChatGPT on a 2026 number JetBrains did not print. The title is about coding agents at work as they cut the subset.
Stay inside the cut.
Self-report can inflate the cool tool and hide the unfashionable one. Weighting for JetBrains familiarity is meant to shrink that. The 2025 write-up still admits leftover bias.
Hobbyists and students are out of the professional-only model. Country-count forecasts on the data playground miss by about 14% MAPE, which is a different pipeline, and a reminder that globally representative is a method, not a census.
31% most-used means 69% have a different most-used tool. US 47% usage is not US most-used. That exclusive split was not published.
Do not invent it. Enterprise policy can still be Copilot-only while the public survey looks like a Claude Code town. Those people are in the 21%, or they are not in the sample as using Claude Code at work.
Either way, your legal review is not a chart. No 2026 raw file to rerun the tabs. JetBrains promised anonymized data when the survey launched. It is not on the methodology page yet.
Anything that needs a crosstab by company size for May to July is a wait.
Standardize on conversion, keep a second agent#
What should you actually standardize on this quarter? Conversion is the market signal, not the whole decision. Awareness is not the signal. The IT SKU is not the signal.
Claude Code currently wins that test among professional developers in this survey. 39% use it at work. 31% live in it.
Copilot still ties the mindshare and still owns a lot of the paperwork. Codex is the chase. OpenCode is the leak if you want an open harness. Cursor is famous and emptying.
If you are writing the standard, treat Claude Code as the current default agent, then still check policy, cost, quota, and review load before you carve it into the repo.
- Name Claude Code as the default agent
- Keep Copilot for autocomplete if people still want Tab
- Keep a second agent ready, because last year's leader was at 3% the year before that
JetBrains' own chart is the argument against a forever lock. If policy is Copilot-only, stop pretending the survey is your building. Budget the switch for the day the vendor list moves.
The HN shops already described that lag. They were not confused. They were stuck.
Review is still the queue that fills when generation gets cheap. That is the bottleneck post, not this one. Adoption is settled enough to pick a default.
The work after the pick is the same as it was when Copilot was the name on the door.
Workplace default questions
Is GitHub Copilot still relevant in the enterprise?
As a license, yes. As the lived-in agent, less so. JetBrains measured Copilot at 21% work usage in May to Jul 2026, down from 29% a year earlier, while Claude Code sat at 39%. January Pulse still had Copilot at 40% inside companies with more than 5,000 people. The SKU can lag the chair by a year.
asked on news.ycombinator.com ↗What if Copilot is the only AI your company allows?
Then the survey is describing a different building than yours. The 39% and 31% numbers are what people say they use at work, not what procurement bought. One HN shop in that thread still had Copilot as the only allowed tool, and the rollout lag for Claude Code was the complaint. Standardize later. Budget a second agent the day the policy moves.
asked on news.ycombinator.com ↗Did shops already replace Copilot with Claude Code?
Some did. In the same thread, one team started on Copilot and was largely replaced by Claude Code because the agent could run scripts. That is a conversion story in one org, matching the survey's 80% conversion among Claude Code users. It is not a mandate to rip out the Copilot seat if it is still doing cheap autocomplete.
asked on news.ycombinator.com ↗