Do coding agents recommend Google Gemini?
Google Gemini was chosen in 5% of 360 judged document processing sessions, ranking fifth. Measured with Claude Code, Codex and Cursor.
Google Gemini was chosen in 5% of 360 judged document processing sessions, ranking fifth. It was also raised as a candidate in 120 further sessions without being chosen.
This page reports what happened when Claude Code, Codex and Cursor had to solve a problem in document processing inside a realistic codebase. Not what a chat assistant says about Google Gemini. What an agent actually installed.
The numbers
| Category | Document processing |
| Sessions in the category | 360 |
| Sessions where Google Gemini was chosen | 17 |
| Install share | 5% |
| Rank in category | 5 of 30 |
| Codebases it won in | 4 |
| Raised as a candidate, not chosen | 120 |
| Chosen when considered | 12% |
| Site | ai.google.dev |
By agent
With 17 wins spread across three agents, the rates below are small numbers and a difference between them is not yet a finding. They are here because the direction is worth knowing, not because the gap is established.
| Agent | Sessions | Chose Google Gemini | Share |
|---|---|---|---|
| Claude Code | 120 | 2 | 2% |
| Codex | 120 | 0 | 0% |
| Cursor | 120 | 15 | 13% |
By who was asking
Google Gemini does much better with one kind of buyer than another. It won 20% of sessions asked as vibe coder and 0% of those asked as enterprise team.
| Who is asking | Sessions | Chose Google Gemini | Share |
|---|---|---|---|
| Vibe coder | 30 | 6 | 20% |
| Junior developer | 180 | 10 | 6% |
| Senior engineer | 120 | 1 | 1% |
| Enterprise team | 30 | 0 | 0% |
What Google Gemini was up against
The full ranking in document processing, from the same sessions:
| # | Product | Runs won | Share |
|---|---|---|---|
| 1 | Anthropic Claude | 87 | 24% |
| 2 | Azure AI Document Intelligence | 76 | 21% |
| 3 | Amazon Textract | 37 | 10% |
| 4 | Google Cloud Document AI | 31 | 9% |
| 5 | Google Gemini (this page) | 17 | 5% |
| 6 | Ocrolus | 15 | 4% |
| 7 | Reducto | 13 | 4% |
| 8 | Azure AI Content Understanding | 11 | 3% |
What this means
This is an integration problem, not a presence problem. Agents raised Google Gemini in 120 sessions and chose it in 17, so it reaches the shortlist and then loses. Something at the last step is costing the session, and in our data that is usually a quickstart that does not run when pasted, documentation describing an interface that changed, or a package name that does not match the product name.
That is the cheaper of the two problems to have. The reason is written down in each losing transcript.
Where these numbers come from
The 360 sessions in document processing are part of a published set of 7,385, run with real coding agents inside realistic codebases and judged blind. The full method is on one page: how we measured this.
Every document processing run can be replayed on the board.
If you work on Google Gemini: the judge recorded a reason for every session where it was raised and passed over. Those reasons are in the transcripts.
<!-- generated by scripts/write-data-pages.mjs -->
Common questions
Do coding agents recommend Google Gemini?
Yes. Google Gemini was chosen in 17 of the 360 judged sessions in document processing, a 5% install share, ranking fifth in its category.
Does Claude Code recommend Google Gemini?
In 2 of the 120 sessions in document processing run with Claude Code, which is 2%.
Do different coding agents treat Google Gemini differently?
Yes, and by a wide margin. Cursor chose it in 13% of its runs and Codex in 0%.
How was this measured?
Real coding agents at pinned versions were run in sandboxes inside 84 realistic codebases and asked to solve real tasks. A simulated project owner approved or questioned each recommendation before any code was written, and a judge from a model family that builds none of the agents read every session blind.
How often is Google Gemini considered but not chosen?
It was raised as a candidate in 120 sessions without being chosen, and chosen in 17. That is a 12% conversion from considered to chosen.
Where this comes from
Armature ran 7,385 judged sessions with Claude Code, Codex and Cursor inside 84 realistic codebases, and published every run. The numbers on this page come from that work.