AI Coding Assistant Statistics: Use vs. Trust
66% of developers use an AI coding assistant, but only 48% trust its output even when they can check it. The open position is the check, not the assistant.
The signal
- LaunchesClaude Opus 5.5 (22 Sept 2026), is Anthropic's launch number 4 in our feed since July.
- LaunchesOpenAI listed GPT-6 Sol and GPT-6 Luna the same day (22 Sept 2026, per launch trackers; the official post was not fetched).
- ArenaOn the 2 Oct 2026 Arena coding board, Google models hold the top 1 places. The next model, mimo-v2.6-pro (Xiaomi), has a rating interval that overlaps theirs, so the edge of that lead is not settled.
66% of developers use AI coding assistants, but only 48% trust AI output they can validate, a gap of 18 points. Agents reach 26%.
Share of respondents, %, Stack Overflow Developer Survey 2026 (AI section, n ≈ 13,000). Daily use is among developers who use AI coding tools; every other bar is among all respondents.
Some toplines were read through search summaries; spot-check against the survey site before publishing. Daily use is among AI-tool users only. Field dates are not stated in the pages we read (a seven-week survey in 2026). Bars are separate questions, so they do not add up. The trust item is conditional ("when they can validate answers"), so it is a generous measure of trust.
Show the data (6 rows)
| Survey item | Share of developers |
|---|---|
| Use AI daily (AI tool users only) | 73 |
| Use AI coding assistants | 66 |
| Use general AI chat tools | 63 |
| View AI at least somewhat positively | 62 |
| Trust AI output they can validate | 48 |
| Use AI agents or workflows | 26 |
The Mode read
Anthropic shipped Claude Opus 5.5 on 22 Sept 2026, and OpenAI listed GPT-6 Sol and Luna the same day. On the Arena coding board, Google models hold the top 1 places. The Stack Overflow Developer Survey 2026 shows where developers stand: they use these tools, but do they believe what the tools write?
Use is high. 66% of respondents use AI coding assistants, 63% use general chat tools, and 73% of AI coding tool users use them every day. Trust is not. Only 48% trust AI output even when they can validate it, 18 points below use. Agents, which act with less supervision, reach 26%, 40 points behind assistants. 62% still view AI at least somewhat positively, so this is not rejection. It is supervision.
Use without trust means every AI-written line still needs a person or a test to approve it. That review step is a distribution point. It sits in the pull request and the CI pipeline, and it sees every change, whichever assistant wrote it. A new model does not remove it, because a better assistant writes more code that still needs checking. The assistant is a crowded race; the checkpoint is a seat each team installs once.
Developers adopted the assistant. The open job is making its code safe to merge.
What we would back: Review, testing and verification tools that sit in the pull request and check every AI-written change, whichever assistant wrote it.