AI coding assistants in 2026: what was measured and what was invented

This is the first thing to know about the subject. The 2026 questionnaire opened on 23 June 2026 and there are no results: survey.stackoverflow.co/2026/ returns 404, the survey home page lists years from 2024 back to 2011, and the internet archive has never captured a 2026 page.

Meanwhile "Stack Overflow 2026" figures circulate through roundups — those are almost always the 2025 numbers, fielded between 29 May and 23 June 2025. The well-known 84% comes from there.

The pattern repeats elsewhere: there is no 2026 Octoverse either (the last edition is 28 October 2025, with data closed on 31 August 2025), and JetBrains has published no complete 2026 report — only fragments on its blog.

Stack Overflow, fielded 29 May – 23 June 2025, 49,009 responses from 177 countries:

The gap between 84% using and 3.1% trusting is the most substantive thing in this data. Almost everybody uses the tool and almost nobody believes it; 66% name the near-correct answer as their chief frustration.

Limits: this is a self-selected sample. Stack Overflow states outright that respondents were recruited through its own channels and that "highly-engaged users were more likely to notice the prompts". There is no random sampling frame and no computable response rate. Note the bases too: they differ per question and are always smaller than 49,009.

Trend for "use or plan to use": 70% in 2023, 76% in 2024, 84% in 2025. A caveat on that line: the sample shrank from 65,437 to 49,009 in a year — by a quarter.

JetBrains fielded from May to July 2026 among more than fifteen thousand professional developers and, unlike Stack Overflow, reweights its results by region, employment, language and familiarity with its own products. The result: 90% use AI coding agents at work at least weekly, 68% daily.

But the tool mix turned over within a year:

Caveat: JetBrains’ full 2026 report is not published — the figures appear piecemeal on its blog. And JetBrains runs a second, separate survey ("AI Pulse", January 2026) whose numbers are not comparable line-for-line with these; the two must not be merged.

Everything vendors say about themselves is a company describing its own product without independent audit. Even so, one comparison is telling.

GitHub Copilot is credited with 50 million users (stated on 29 July 2026 in an investor call — the figure appears in no written press release). Meanwhile paid subscribers number 4.7 million (GitHub, 28 January 2026). So roughly nine in ten "users" are not paying: the count spans free, student, open-source and trial tiers, and it is cumulative all-time rather than active users.

Revenue at private companies is quoted as ARR — a short recent period multiplied out to a year, not booked and confirmed money. Anthropic cited $1B for Claude Code in December 2025 and $2.5B in February 2026; Cursor cited over $1B in November 2025. None of them publishes audited financials.

One figure is better not repeated. The claim that "Copilot writes 46% of code" is everywhere, but it traces to no GitHub publication at all. GitHub-funded research discusses suggestion acceptance rates of roughly 21–30% — an entirely different quantity.

Three rules follow directly from the above:

1. Check the collection year, not the publication year. The difference here reaches a year and more: Octoverse appeared in October 2025 with data closed in August 2025; the 2025 Web Almanac was published in January 2026 from a July 2025 crawl.
2. Look at the question base, not the survey size. The survey had 49,009 participants, but 33,244 answered the trust question and 31,877 the agents question. Percentages are computed on the question base.
3. Separate "users", "active" and "paying". The gap between 50M and 4.7M is not a contradiction but two different quantities, and presenting the first as audience size is wrong.

Sources: Stack Overflow 2025, AI section, its methodology, JetBrains, 2026 fielding, GitHub Octoverse.