Claude Code vs Cursor vs GitHub Copilot: What the 2026 Adoption Data Shows

GitHub Copilot has the largest paid base in AI coding and the fastest shrinking share of daily work. JetBrains surveyed more than 15,000 professional developers between May and July 2026 and put Copilot adoption at work at 21%, down from 29%. On the same survey, Claude Code vs Cursor is no longer close: 39% against 12%. The ranking you get depends on which number you agree to count.

The short answer

For most teams in 2026, Claude Code is the strongest default, Cursor is the better buy for developers who live inside one editor, and GitHub Copilot is the safest enterprise purchase. Adoption data favours Claude Code. Distribution still favours Copilot. No benchmark result should decide this for you.

What changed between January and July 2026

The category reordered itself in about seven months. JetBrains ran an AI pulse survey of more than 10,000 developers in January 2026, then fielded its main Developer Ecosystem Survey across more than 15,000 professional developers from May to July.

Between those two waves, Claude Code went from 18% adoption at work to 39%. GitHub Copilot fell from 29% to 21%. Cursor fell from 18% to 12%. OpenAI's Codex went from 3% to 16%.

Two of those moves are large enough to matter and two sit inside the range where sampling noise could explain part of the gap. The Claude Code rise and the Codex rise are the ones I would treat as real. Copilot's decline is directionally clear across both waves. Cursor's is the number I trust least.

The Cursor line is the one worth pausing on. More developers knew about the product by July and fewer were using it at work. That combination usually means trial without retention, or displacement by a tool that arrived later in the same slot.

Copilot leads the installed base, and then the metric changed

GitHub Copilot is still the most widely known product in the category, with 79% global awareness and higher figures across Europe and the US. It also has the deepest enterprise distribution, because it arrives attached to a source control system most engineering organisations already pay for.

26 million users, then 4.7 million subscribers

Microsoft reported 15 million Copilot users in April 2025, 20 million in July 2025 and 26 million in October 2025. At its FY26 Q2 earnings call on 28 January 2026, the disclosure changed shape. Microsoft reported 4.7 million paid subscribers, up about 75% year on year.

Those two numbers measure different things, so the comparison is not a like-for-like decline. It is still informative. A 26 million user base and a 4.7 million paying base implies that roughly 18% of the people who had touched the product were paying for it three months later.

Microsoft does not break out GitHub Copilot revenue. Every ARR figure you see quoted for Copilot is a calculation someone performed on the subscriber count and an assumed price mix. Treat those as estimates, not disclosures. Be suspicious of any comparison table that puts an estimated Copilot number next to a company-reported Cursor number without saying so.

Distribution is still the strongest asset in this category

The most interesting thing GitHub did in 2026 was concede the model layer. In February, GitHub put Anthropic's Claude and OpenAI's Codex agents into public preview on GitHub and VS Code for Copilot Pro+ and Enterprise, then extended it to Business and Pro users three weeks later. No extra subscription is required.

Read that as a strategy, not a surrender. GitHub is betting that the durable position is the place where work is assigned, reviewed and merged, rather than the model that writes the patch. That is the same argument as owning the workflow rather than the intelligence, which we have made before about where integration depth becomes a real moat.

If that bet works, the tool comparison stops mattering, because the agent becomes a dropdown. If it fails, Copilot has trained its own buyers to prefer a competitor's agent inside its own product.

Cursor's growth is real, and its unit economics are split

Cursor is the commercial success story of the category. Its parent, Anysphere, closed a $2.3 billion Series D at a $29.3 billion post-money valuation in November 2025. By April 2026 it was in talks to raise more than $2 billion at a $50 billion pre-money valuation. In June 2026, SpaceX agreed to acquire the company for $60 billion in stock, with the deal expected to close in the third quarter.

Bloomberg reported Cursor at $2 billion in annualised revenue in February 2026, roughly doubling from $1 billion in late November 2025. People familiar with the company told TechCrunch it was projecting more than $6 billion annualised by the end of 2026.

Annualised run rate is not revenue. It takes a recent period and multiplies it out, which flatters any business growing this fast and tells you nothing about retention. We have written about why the fastest revenue ramp in software history is also the hardest one to underwrite.

Individual accounts lose money, enterprise accounts do not

The detail from the April reporting that almost nobody quoted is the one that matters most to a buyer. Cursor's enterprise accounts carry positive gross margins. Its individual developer accounts carry negative gross margins. The company had only recently reached slight gross margin profitability overall.

That asymmetry explains the pricing behaviour across this entire category. A $20 per month plan serving an agentic workload is a loss leader whenever the user is heavy, and the vendor's only routes out are usage caps, credit systems or a price rise. This is the same dynamic driving the shift from seats to credits and tokens, and it is why inference cost sits directly on the gross margin line for every vendor here.

My position: if you are buying individual plans for a team of heavy users, you are buying something the vendor needs to reprice. Budget for that rather than for the sticker.

Prices were read from the vendors' public pricing pages on the date above. Every one of these vendors has changed limits or credit allowances at least once in the past year, so verify before you budget. Additional GitHub AI Credits were listed at $0.01 each.

Claude Code leads satisfaction, and that is a softer signal than it looks

In the January 2026 pulse, JetBrains recorded a customer satisfaction score of 91% and a net promoter score of +54 for Claude Code, the highest of any tool it measured. By the May to July wave, Claude Code was the single most used tool for 31% of developers.

That second figure is the one I would act on. It implies close to an 80% conversion from regular use to primary tool. Satisfaction scores tell you how people feel about a product they already chose. Conversion to primary tool tells you what happens when a developer has three options open and picks one.

Why satisfaction runs ahead of adoption here

Claude Code arrived later and grew from a small base, and new tools in any category poll well before the awkward parts surface. Awareness sat at 31% in mid-2025 and 57% by January 2026. Products measured during a steep adoption curve are measured on their enthusiasts.

The honest read is that the satisfaction lead is real and probably overstated. I would discount it by a wide margin and still conclude Claude Code is ahead, because the primary-tool conversion number holds up independently of how anyone feels.

Why benchmark scores cannot settle this

Every vendor in this category quotes a coding benchmark. The most cited is SWE-bench Verified, and its composition does not resemble the work your team does.

The benchmark is mostly short bug fixes in a handful of repositories

Epoch AI's analysis of what SWE-bench Verified actually measures found 500 Python problems drawn from 12 open source projects, with Django accounting for close to half. About 39% of tasks are rated as trivial changes under 15 minutes and 52% as small changes under an hour. Roughly 87% are bug fixes. Models change an average of 1.87 functions per task.

Epoch also estimated that 5% to 10% of the tasks are flawed or unsolvable, that five repositories account for more than 80% of samples, and that scaffolding choices alone can move a score by up to 20 percentage points. About half the issues date from 2020 or earlier, which means the code is almost certainly in the training data. OpenAI has published its own reasoning for no longer evaluating on SWE-bench Verified.

A tool that is excellent at 40 minute Python bug fixes in Django may be poor at a six hour refactor across a TypeScript monorepo with a flaky test suite. The benchmark cannot tell you which one you are buying.

What the field evidence says instead

The most careful field measurement remains METR's randomised controlled trial from July 2025. It found that experienced open source developers took 19% longer to complete real tasks when allowed to use AI tools, while estimating afterwards that AI had made them 20% faster.

Read the caveats before you use that number as a weapon. The study covered 16 developers and 246 tasks in mature repositories they knew well, using tools available between February and June 2025, primarily Cursor Pro with Claude 3.5 and 3.7 Sonnet. METR itself now labels the result as historical. It is one careful study, not a settled finding.

Google's 2025 DORA research points the same way on a different axis. Its report on AI-assisted software development found AI adoption positively associated with delivery throughput and negatively associated with delivery stability. AI amplifies whatever practice already exists. Teams with weak tests and slow feedback loops get faster at shipping problems.

Running two or three tools in parallel is the default, not a failure

By mid-2026, 90% of surveyed developers used an AI coding agent weekly and 68% used one daily. Those figures sit alongside tool-level adoption numbers that sum well past 100%, which is only possible if a large share of developers run more than one.

The common stack is an editor-integrated assistant for in-flow completion, an agent for multi-step work, and a chatbot for reasoning about a problem before touching code. Copilot is strong in the first slot, Claude Code in the second, and the third is contested.

What the duplication actually costs

Two subscriptions per developer at $20 to $40 each is $480 to $960 per developer per year before usage overage. For 40 engineers that is a $19,000 to $38,000 line item that nobody approved as a single decision.

The larger cost is not the licence. It is the review burden. Stack Overflow's 2025 survey found that the top developer frustration was AI output that is almost right but not quite, cited by 66%, with 45.2% saying debugging AI-generated code takes more time. GitClear's research on its own corpus of changed code points the same way, with duplicated blocks rising and refactoring falling as AI authorship grew.

I would not run a three tool stack on purpose. I would pick a primary, allow one secondary, and put the third on an expensed personal plan rather than a company contract. That is a governance decision, and it belongs in the same conversation as the wider problem of tool sprawl inside software teams.

Where this comparison is weakest

Everything above rests on survey data from one research team plus company disclosures of differing quality. Here is what would change the answer.

Survey panels are not the market

The adoption figures come from JetBrains, which sells competing IDE and AI products and recruits partly from its own user base. It reweights by region, employment status, language and JetBrains familiarity, which is more disclosure than most vendors provide. That does not eliminate the problem. A panel skewed toward JetBrains IDE users will under-represent developers who live in VS Code, and VS Code is where Copilot is strongest.

The Stack Overflow trust figures carry the same shape of bias in the other direction. Stack Overflow's audience has a structural reason to be sceptical of tools that reduce traffic to Stack Overflow.

The case for staying on Copilot is stronger than the trend suggests

If you already pay for GitHub Enterprise, Copilot is a line item extension rather than a new vendor. Your security review is done. Your data residency question is answered. Your admin controls exist. And since February 2026 the product will run Claude and Codex agents for you anyway.

Against that backdrop, a 21% adoption figure is not a reason to migrate. I disagree with the common reading that Copilot is finished. What the data supports is that Copilot has lost the argument about which agent is best, not the argument about where agents should live. Those are different fights and it is currently winning the second one.

The genuinely unresolved question is whether the model layer commoditises. If it does, distribution wins and this whole comparison ages badly within a year. That is the same question behind the push to standardise how agents talk to tools, and behind the build versus buy decision for coding agents.

How to choose, in four questions

Skip the feature matrix. Four questions decide this in practice, and the honest answer to each includes where the other option wins.

The fourth row is the one most buying committees skip, and it has the largest effect on the outcome. Throughput rises with AI adoption and stability falls. If a regression currently takes your team three days to find, faster code generation makes that worse, not better.

Frequently asked questions

Where to start this week

Two things, and neither requires a vendor call.

First, find out what your team already runs. Pull the expense reports and the SSO logs rather than asking, because the answer from a survey and the answer from billing are rarely the same. You are looking for how many distinct AI coding tools you already pay for, and how many of those seats went unused last month.

Second, pick one recent piece of work that took longer than it should have and run it through two of these tools yourself. Not a benchmark task. The actual ticket, in the actual repository, with the actual test suite. That single exercise will tell you more than any comparison table, including this one.

Related reading