From Aryan Vatsa | Product & Market Analysis
GitHub Copilot's Market Share Did Fall. The 51% Figure Is Not the Proof
On this page
A number doing the rounds says GitHub Copilot fell from 67% to 51% of developers. The 67% is real, and much narrower than it sounds. The 51% is not in the survey that every GitHub Copilot market share post cites. The decline itself is real anyway, and the honest figure is 21%.
Key takeaways
- The 67% is a multi-select recall question answered by 8,323 people. Stack Overflow's 2025 survey asked which out-of-the-box agents and assistants respondents had used. Copilot scored 67.9% among that group, which is 17% of the survey's respondents, not 67% of developers.
- The 51% does not appear in the published source. There is no 2026 Stack Overflow Developer Survey. The most recent published edition is the 2025 one, released on 29 December 2025, and no Copilot figure in it reads 51%.
- The real fall is 29% to 21%, in JetBrains data. Copilot's regular use at work dropped across two survey waves, while Claude Code went from 18% to 39% and Cursor fell from 18% to 12%.
- Copilot's revenue rose while its share fell. Microsoft reported 50 million Copilot users and Copilot revenue accelerating over 60% quarter over quarter on its 29 July 2026 earnings call. Contracts and preference are now moving in opposite directions.
Where the 51% figure came from, and what it is not
The claim travels in a fixed shape. Copilot fell from 67% to 51% of developer share, usually credited to the Stack Overflow Developer Survey, usually with no link. The first half of that sentence checks out. The second half does not survive an afternoon.
What the survey actually asked
The 67% comes from the 2025 Stack Overflow Developer Survey, published on 29 December 2025. The question was whether respondents had used any of a list of out-of-the-box agents, copilots or assistants.
Copilot placed second on 67.9%, behind ChatGPT on 81.7%. Google Gemini took 47.4% and Claude Code 40.8%. Read the question again and the ranking makes sense. It is a recall list, and the tools everyone has been given by default sit at the top.
That is not market share. It is ever-used, multiple answers allowed, in a survey where nobody was asked to name a primary tool.
Why the base of 8,323 matters
That question drew 8,323 responses, which the survey reports as 17% of respondents. The AI usage question in the same section drew 33,662.
So the 67.9% describes a self-selected sixth of the sample, answering a question about agents. Compare that with the survey's core AI question, where 47.1% said they use AI tools daily. Different bases, different questions, different meanings.
A number quoted without its base is a number you cannot use. This one gets quoted without its base almost every time.
Six restatements are not corroboration
No figure of 51% for Copilot appears in the published 2025 cut. There is also no 2026 Stack Overflow Developer Survey to have produced one. Stack Overflow's own survey index lists the 2025 edition as the most recent, dated 29 December 2025.
My best guess is a collision with a different statistic. Coverage of the same survey carries a 51% figure attached to daily AI tool use among professional developers, and the published cut shows 47.1% using AI tools daily across all respondents. I cannot prove that is the origin. Neither can the posts repeating it, which is the actual problem.
I would not publish that number, and I would not sign a renewal on the strength of it. When six posts carry the same striking figure and none links a source, that is one unsourced claim with five echoes.
The decline that is real is measured somewhere else
None of this means Copilot is fine. It means the evidence for the decline is in a different dataset, and it is stronger than the one being quoted.
29% to 21% across two waves
JetBrains runs a developer survey with a consistent question about regular tool use at work. In the January 2026 wave of more than 10,000 professional developers, Copilot led on 29%. JetBrains noted then that its growth in both awareness and adoption had stalled.
By the May to July 2026 wave of more than 15,000 developers, Copilot was at 21%. Same publisher, same question, two readings, 8 points apart. That is a measured decline with a named base, which is exactly what the 51% claim lacks.
Claude Code is where the share went
Claude Code moved from 18% to 39% across the same two waves, and 47% in the United States. Codex went from 3% to 16%. Cursor fell from 18% to 12%, which is the detail most coverage drops, because it breaks the tidy story of one challenger beating one incumbent.
Two things happened at once. Copilot lost ground, and the entire completion-first category lost ground to agent-first tools. Cursor's revenue kept climbing through the same period, which is the subject of the piece on how fast that business scaled, and it still lost survey share. Revenue and preference are not the same instrument.
| Source and date | Question | Copilot figure and base |
|---|---|---|
| Stack Overflow 2025, published 29 December 2025 | Which out-of-the-box agents, copilots or assistants have you used | 67.9% of 8,323 responses, 17% of respondents |
| JetBrains, January 2026 wave | Which AI tools do you regularly use at work | 29% of more than 10,000 professional developers |
| JetBrains, May to July 2026 wave | Which AI tools do you regularly use at work | 21% of more than 15,000 professional developers |
| The circulating claim | Not stated | 51%, no base, no linked question |
Rows one and two are not comparable and were never meant to be. Ever-used recall and regular use at work are different measurements, which is why a decline calculated across them is meaningless.
Awareness is not adoption, and Copilot's gap is the widest
The most useful number in the JetBrains data is not the drop. It is the distance between how many developers know a tool and how many use it.
Copilot is known to 79% of developers globally, and JetBrains reports awareness between 86% and 90% in Europe, the United Kingdom and the United States. Adoption is 21%. That gap of 58 points is the largest in the category.
Cursor sits at 75% awareness and 12% adoption. Codex is 65% and 16%. OpenCode, the open-source agent, is 42% and 7%, so the newest entrant already converts awareness at a higher rate than the incumbent does.
Distribution built that 79%. GitHub ships to almost every professional developer, and Microsoft sells into almost every large enterprise. What distribution cannot do is hold usage once a developer has tried something better, a dynamic covered from the other direction in the piece on when the worse product wins.
Why developers switched, in the order it happened
Share moves for reasons you can date. Three of them landed inside six months, and only one of them is about model quality.
The June 2026 billing change
On 1 June 2026, GitHub moved every Copilot plan to usage-based billing. Premium request units were replaced by GitHub AI Credits, metered on input, output and cached tokens at published API rates. Pro+ stayed at $39 a month and now includes $39 of credits. Business is $19 per user with $19 of credits.
The list price did not rise. The certainty did fall. On GitHub's own community forum, developers reported single agent runs consuming hundreds of credits, and one Pro+ subscriber calculated their monthly allowance would be gone in under two days. The Register reported on 2 June 2026 that users were threatening to leave, with one describing $6 of credits consumed by a single request as impossible to predict.
GitHub's defence is sound on the economics. A quick chat question and a multi-hour autonomous session used to cost a subscriber the same amount, which is not a pricing model, it is a subsidy. The mechanics of that shift, and why credits confuse buyers so reliably, are set out in the breakdown of credit and token billing.
What the change also did was end the reason many teams stayed. A predictable seat fee was Copilot's quiet advantage over metered rivals. Once every option bills on consumption, the comparison collapses to output quality, and that is a comparison Copilot has been losing. Related pressure on seat-based models is covered in the analysis of seat compression.
Reliability showed up on the agent path
Agentic coding fails differently from autocomplete. A dropped completion is an annoyance. A stalled agent job is a blocked task with context lost.
GitHub's own incident record for 2026 includes Copilot degradations of exactly that type. On 5 August 2026, an internal rate limit applied more broadly than intended delayed 100% of newly submitted cloud agent jobs for about 52 minutes. On 17 August 2026, a wider GitHub incident ran for close to eight hours, and Copilot stayed impaired after other services were declared operational.
One outage changes nothing. A pattern of them, during the exact months a developer is deciding which tool to open by default, changes plenty.
Suggestion quality is the claim nobody can measure
The community complaint that Copilot's suggestions got worse is the least verifiable part of this story, and I am not going to pretend otherwise. There is no public, comparable benchmark of acceptance rates across vendors.
The nearest public proxy is satisfaction. JetBrains reported Claude Code at 91% satisfaction and a net promoter score of 54 in the January 2026 wave. No equivalent figure was published for Copilot, so the comparison is one-sided by default. Treat it as directional and unresolved.
The contract base is still holding, and that is the real story
Here is where the decline narrative runs into Microsoft's numbers.
On the FY26 Q4 earnings call on 29 July 2026, Satya Nadella said GitHub Copilot has 50 million users, that one in three pull requests on GitHub now involves an agent, and that more than 90% of the Fortune 500 use GitHub for AI-powered development. Amy Hood attributed stronger-than-expected Copilot consumption to the June billing change.
Copilot revenue accelerated over 60% quarter over quarter in the same quarter its survey share fell. Both facts are true, and reconciling them is the whole point.
Metering explains most of it. When you bill on consumption rather than seats, revenue tracks the tokens burned by the developers who stayed, not the number of developers who prefer you. A shrinking, heavier-using base can produce a rising revenue line for several quarters.
Enterprise procurement explains the rest. Copilot is bundled into agreements that already cover identity, repositories, CI and cloud. Nobody unpicks that in a quarter because a survey moved. The switching cost is contractual, not technical, and it is the same dynamic examined in the build versus buy analysis for coding agents.
Where this argument is weakest
Four objections, stated as strongly as I can put them.
The JetBrains base is not neutral. JetBrains sells IDEs and an AI assistant that appears in the same rankings. Its panel skews toward its own users, and any vendor-run survey carries house effects. I use it because the question wording and base are published and consistent across waves, not because it is disinterested.
Adoption questions measure recall, not telemetry. Nobody in either survey was instrumented. Developers report what they remember using, and the tool they talk about is not always the tool running in the background. Copilot's inline completions are exactly the kind of usage that goes unreported.
JetBrains dates its own comparison loosely. The August 2026 post describes the fall as 29% a year ago to 21% now, while the April 2026 post puts 29% in the January 2026 wave. Either the figure sat flat for months or the comparison period is imprecise. I could not reconcile the two statements from the published posts, so both are reported here rather than one being quietly picked.
Share of a survey may be the wrong metric entirely. If you hold Microsoft stock, 21% is close to irrelevant next to 50 million users and accelerating consumption revenue. The case for caring about developer preference is that it leads renewals by 12 to 24 months. That is a belief about lag, not a proven relationship, and I hold it loosely.
What the gap between contracts and preference costs you
If you buy tools for a team, the interesting question is not who leads a survey. It is whether you are paying for a tool your developers have quietly stopped opening.
That gap has a specific cost now that billing is metered. Unused seats used to be waste with a fixed ceiling. Metered plans convert them into an allowance you paid for and did not consume, sitting next to a second tool your team expensed on a personal card.
| Check | How to run it | What a bad reading looks like |
|---|---|---|
| Credit consumption per seat | Pull the billing dashboard and rank seats by credits used in the last full month | A long tail of seats using a small fraction of the included allowance |
| Shadow spend on rival tools | Search expense claims and card statements for AI coding tools bought outside procurement | Developers paying personally for the tool they actually work in |
| Agent path reliability | Log every blocked or restarted agent run for four weeks, with the time lost | Repeat incidents on the same workflow, absorbed silently by the team |
The second row usually produces the finding. Shadow spend is the closest thing a buyer has to a preference survey with a real base, and it costs nothing to run.
My own position, stated plainly. I would not switch a large team off Copilot on the strength of survey movement, because the integration surface is worth more than the eight points. I would renew for one year rather than three, and I would set a spending limit before June's meter teaches me what a bad month costs.
Frequently asked questions
Did GitHub Copilot's market share really fall from 67% to 51%?
No source supports that pair of numbers. The 67% is from Stack Overflow's 2025 survey, where 67.9% of 8,323 respondents said they had used Copilot at some point. No published Stack Overflow figure reads 51% for Copilot, and there is no 2026 edition of that survey. The decline is real, but it is documented in JetBrains data at 29% to 21%.
What is GitHub Copilot's market share in 2026?
It depends entirely on the question asked. JetBrains, surveying more than 15,000 professional developers between May and July 2026, put regular use of Copilot at work at 21%, behind Claude Code at 39% and ahead of Codex at 16% and Cursor at 12%. Microsoft separately reports 50 million Copilot users, which measures footprint rather than preference.
Why are developers switching from GitHub Copilot?
Three things arrived together. GitHub moved all plans to usage-based billing on 1 June 2026, which removed the predictable seat price that kept many teams in place. Copilot suffered repeated agent-path incidents through 2026, including a full delay of new cloud agent jobs on 5 August. And agent-first rivals scored better on published satisfaction measures.
Is GitHub Copilot losing to Claude Code and Cursor?
It is losing to Claude Code, which rose from 18% to 39% adoption at work across two JetBrains waves. It is not losing to Cursor, which fell from 18% to 12% in the same period. The shift is from completion-first tools toward agent-first ones, and Copilot and Cursor both sit on the wrong side of it.
How much does GitHub Copilot cost after the 2026 billing change?
Headline prices held. Pro is $10 a month including $10 of AI Credits, Pro+ is $39 including $39 of credits, Business is $19 per user and Enterprise is $39 per user, each including credits equal to the seat price. Credits are consumed on input, output and cached tokens at published model rates, so heavy agent use bills above the seat fee.
Should I cancel GitHub Copilot at renewal?
Not on survey data alone. Run three checks first: credit consumption per seat from the billing dashboard, expense claims for rival tools bought outside procurement, and a four-week log of blocked agent runs. If most seats burn a fraction of their allowance while developers expense something else personally, you have a preference problem that a shorter renewal term addresses better than a migration.
Where to start this week
Start with the number in your own account, not the one in the headline. Open the Copilot billing dashboard, sort seats by credits consumed last month, and read the bottom quartile. That list is your real adoption rate, measured rather than surveyed.
Then do one thing with the 51% claim itself. Whoever sent it to you, ask them for the question wording and the response count behind it. That request takes a sentence, and it is the fastest way to find out whether the rest of the analysis you were sent was built the same way.
Related analysis
Copilot's metered turn is one instance of a wider repricing. Read how per-seat pricing is being taken apart, and what agent interoperability standards do to switching costs once tools stop owning the workflow.
References
- Stack Overflow, 2025 Developer Survey, AI section, published 29 December 2025. Used for the 67.9% Copilot figure, the 8,323 response base and the 47.1% daily usage figure.
- JetBrains, AI Coding Agents: Adoption Trends, August 2026. Used for the May to July 2026 adoption and awareness figures across all tools.
- JetBrains, Which AI Coding Tools Do Developers Actually Use at Work, April 2026. Used for the January 2026 wave figures and the Claude Code satisfaction scores.
- Microsoft, FY26 Q4 earnings call materials, 29 July 2026. Used for the 50 million users, the pull request figure and the Copilot revenue growth statement.
- GitHub, GitHub Copilot is moving to usage-based billing, 2026. Used for the credit allotments, plan prices and the token metering method.
- The Register, Angry devs vow to flee GitHub Copilot as metered billing takes hold, 2 June 2026. Used for the developer reaction and the reported credit burn examples.
The weakest thing about this source base: both surveys are self-reported and neither is instrumented, and the stronger of the two is run by a vendor that competes in the category it measures. The GitHub incident details are drawn from GitHub's published status history, which records duration and cause but not the number of users affected.
Related reading