← All posts
  • #statistics
  • #ai terminal
  • #claude code
  • #adoption
  • #research

AI Terminal Statistics 2026: What Developers Actually Run

· · 6 min read

Adoption numbers, trust numbers, and repo growth for AI coding tools in 2026, with sources named and estimates labeled as estimates.

Most "AI coding statistics" posts are a pile of numbers with no source and no date. This one names where each figure comes from, says plainly when something is an estimate, and shows you how to check the live ones yourself.

Ground rule I set for myself: if I could not point at a public source, I left the number out. There are a few places below where the honest answer is "nobody has published this," and I say so instead of inventing a percentage.

Adoption: the one number everyone quotes

The most cited figure comes from the Stack Overflow Developer Survey, which is the largest public developer survey and runs annually with tens of thousands of respondents.

The 2025 edition found roughly 84% of respondents using or planning to use AI tools in their development work, up from about 76% the year before. Actual current use was lower than that headline, in the neighborhood of 50 to 60% depending on how you read the question, because "planning to use" carries a lot of the total.

The more interesting finding was the sentiment reversal. Favorable sentiment toward AI tools fell year over year, to roughly 60% from above 70% previously. And on accuracy, more respondents reported distrust than trust: roughly 46% distrusting AI output accuracy against about 33% trusting it.

Read that pair together, because it is the actual story of the year. Adoption up, trust down. People are using these tools daily and reporting that they do not fully believe the output. That is not a contradiction, it is what mature tool use looks like.

Caveat worth stating: the Stack Overflow survey is self-selected. Its respondents skew toward people who use Stack Overflow, which is a specific population. Treat it as directional, not as a census.

Check whether the 2026 edition has landed before citing the 2025 numbers as current. It usually publishes mid-year.

Repo growth: the trajectory story

GitHub stars are a bad proxy for usage and a decent proxy for attention. I am giving orders of magnitude rather than precise counts, because star counts move daily and a precise number in a blog post is wrong within a week.

As of the first half of 2026, roughly:

  • aider, the longest-running open source terminal coding agent, sits in the mid tens of thousands of stars. It has been growing steadily since 2023 rather than spiking.
  • ghostty, the GPU-accelerated terminal, went public in December 2024 and took tens of thousands of stars within weeks. One of the fastest terminal launches on record.
  • opencode grew very fast through 2025 after launch and is now in the same tens-of-thousands band. Its curve is the steep one.
  • Warp's main repository is issue tracking rather than source, since the app is closed. Its star count reflects interest, not code.

Do not take my word for any of these. One command gets you today's number:

curl -s https://api.github.com/repos/Aider-AI/aider | \
  python3 -c "import sys,json; d=json.load(sys.stdin); print(d['stargazers_count'])"

Swap the repo path for whichever project you care about. That is the citation I would rather give you than a stale integer.

The shape matters more than the level anyway. Terminal-native AI tools went from a curiosity in 2023 to multiple projects with five-figure star counts by 2026, and the newest entrants climbed fastest. Attention is compounding in this category, not flattening.

npm downloads: the closest thing to real usage

The major agent CLIs ship as npm packages, and npm publishes download counts. This is the best public usage signal available, and you can query it directly:

curl -s "https://api.npmjs.org/downloads/point/last-week/@anthropic-ai/claude-code" 
curl -s "https://api.npmjs.org/downloads/point/last-week/@openai/codex"
curl -s "https://api.npmjs.org/downloads/point/last-week/@google/gemini-cli"

I am deliberately not printing the numbers I got, because weekly download counts move a lot and a snapshot in a blog post is misinformation within a month. Run the commands. It takes ten seconds and the answer is current.

Two things to know when you interpret whatever you get back. Download counts include CI systems and mirrors, so they overstate humans. And they undercount anyone who installed through a native installer or a package manager other than npm, which is a growing share. The number is a floor with noise on top, not a user count.

Revenue and company figures

This is where most stats posts start fabricating, so let me be direct about what is and is not known.

Anthropic has publicly discussed Claude Code's growth, and business press through 2025 reported revenue run-rate figures in the hundreds of millions of dollars annualized. Those are press-reported figures, not audited disclosures, and the reporting varied. Treat them as evidence that the product got very large very quickly, not as an accounting fact.

OpenAI open sourced Codex CLI and has discussed adoption in launch materials and developer events. Google ships Gemini CLI with a free tier and has published usage limits, which tells you something about intended scale but not about actual users.

What nobody has published, as far as I can find: reliable per-tool active user counts, retention numbers, or a credible market share split between Claude Code, Codex, Gemini CLI, and the open source agents. If you see a pie chart claiming those percentages, ask where it came from. Usually the answer is a survey of a few hundred people or nothing at all.

What I cannot source, and won't pretend to
  • How many developers run more than one agent at a time. Anecdotally common. No public data.
  • Median number of concurrent sessions. No public data. Our own users skew high, which makes our numbers useless as a general statistic, so I am not quoting them as one.
  • Terminal market share among developers. Fragmented, platform-dependent, and mostly guessed at in every post that claims it. See what terminals developers use for the honest version of that question.
  • Token spend per developer per month. Varies by two orders of magnitude between users. An average would mislead.
What the numbers actually say

Three conclusions I think the public data supports.

Adoption is close to saturated in terms of trial. When 84% of a large developer survey is using or planning to use AI tools, the growth question is no longer "will people try this." It is retention and depth of use.

Trust is the real constraint. Distrust in accuracy outrunning trust, in the same year adoption climbed, is the most important finding in the survey. It explains why review workflows, diff quality, and agent transparency matter more to daily users than raw capability benchmarks do.

The terminal is where the serious usage went. Every major lab shipped a CLI, and the open source terminal agents grew fastest of anything in the category. That is a strong signal about where people who use these tools all day actually work. More on that shift in the terminal is getting an AI makeover.

Using these numbers

Cite them with dates. Link the primary source, not this post, when the primary source exists. Re-run the API commands rather than copying a figure, because half of these move weekly.

And if you are one of the people running several agents at once, which the surveys still do not measure: that is the workflow MOLTamp was built for, with a tab per session and badges that tell you which agent is waiting on you. The community marketplace is where people share the setups they actually use.

Corrections welcome. If I got a figure wrong or a better source exists, I would rather fix it than defend it.