Do coding agents recommend OpenAI Realtime API?
OpenAI Realtime API was chosen in 12% of 308 judged voice agents sessions, ranking third. Measured with Claude Code, Codex and Cursor.
OpenAI Realtime API was chosen in 12% of 308 judged voice agents sessions, ranking third. It was also raised as a candidate in 87 further sessions without being chosen.
This page reports what happened when Claude Code, Codex and Cursor had to solve a problem in voice agents inside a realistic codebase. Not what a chat assistant says about OpenAI Realtime API. What an agent actually installed.
The numbers
| Category | Voice agents |
| Sessions in the category | 308 |
| Sessions where OpenAI Realtime API was chosen | 37 |
| Install share | 12% |
| Rank in category | 3 of 18 |
| Codebases it won in | 10 |
| Raised as a candidate, not chosen | 87 |
| Chosen when considered | 30% |
| Site | openai.com |
By agent
The three agents treat OpenAI Realtime API very differently. Codex chose it in 29% of its sessions in this category and Claude Code in 0%, a spread of 29 points.
| Agent | Sessions | Chose OpenAI Realtime API | Share |
|---|---|---|---|
| Claude Code | 94 | 0 | 0% |
| Codex | 110 | 32 | 29% |
| Cursor | 104 | 5 | 5% |
A single blended install share would report the midpoint and hide both ends. If most of your users run Claude Code, your real position here is 0%, not 12%.
By who was asking
OpenAI Realtime API performs similarly across the four kinds of buyer, from 10% to 17%. That is unusual: the category leader changed with the persona in 14 of the 18 categories we measured.
| Who is asking | Sessions | Chose OpenAI Realtime API | Share |
|---|---|---|---|
| Vibe coder | 62 | 6 | 10% |
| Junior developer | 42 | 7 | 17% |
| Senior engineer | 128 | 14 | 11% |
| Enterprise team | 76 | 10 | 13% |
What OpenAI Realtime API was up against
The full ranking in voice agents, from the same sessions:
| # | Product | Runs won | Share |
|---|---|---|---|
| 1 | Vapi | 74 | 24% |
| 2 | Retell AI | 55 | 18% |
| 3 | OpenAI Realtime API (this page) | 37 | 12% |
| 4 | LiveKit Agents | 35 | 11% |
| 5 | ElevenLabs Agents | 30 | 10% |
| 6 | Twilio ConversationRelay | 28 | 9% |
| 7 | Built in-house (no product adopted) | 18 | 6% |
| 8 | Azure Voice Live API | 10 | 3% |
What this means
OpenAI Realtime API was raised in 87 sessions and chosen in 37. That ratio is balanced enough that the ceiling is presence rather than integration: the product converts reasonably when it is on the table, and it is not on the table often enough.
Where these numbers come from
The 308 sessions in voice agents are part of a published set of 5,292, run with real coding agents inside realistic codebases and judged blind. The full method is on one page: how we measured this.
Every voice agents run can be replayed on the board.
If you work on OpenAI Realtime API: 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 OpenAI Realtime API?
Yes. OpenAI Realtime API was chosen in 37 of the 308 judged sessions in voice agents, a 12% install share, ranking third in its category.
Does Claude Code recommend OpenAI Realtime API?
In 0 of the 94 sessions in voice agents run with Claude Code, which is 0%.
Do different coding agents treat OpenAI Realtime API differently?
Yes, and by a wide margin. Codex chose it in 29% of its runs and Claude Code in 0%.
How was this measured?
Real coding agents at pinned versions were run in sandboxes inside 51 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 OpenAI Realtime API considered but not chosen?
It was raised as a candidate in 87 sessions without being chosen, and chosen in 37. That is a 30% conversion from considered to chosen.
Where this comes from
Armature ran 5,292 judged sessions with Claude Code, Codex and Cursor inside 51 realistic codebases, and published every run. The numbers on this page come from that work.