- A speaker at OpenAI's DevDay priced a coding agent at about $32 an hour, later $16, and said fixing CI involves more than the agent.
- Code review and CI resources are extra. A separate aside noted that getting back to green can mean deleting a failing test.
At OpenAI's DevDay 2026, a speaker on the show put an hourly price on a coding agent: about $32, later $16. The same speaker said the figure leaves out code review, the testing machines and other factors. The transcript does not name the speakers.
What was said
A host asked what a six-week run of Codex, OpenAI's coding agent, had really cost. Many people assume OpenAI staff have unlimited tokens, the units AI models are billed in. The speaker said they worked back from the pricing. It came to about $32 an hour to keep Codex fixing CI across OpenAI. CI, or continuous integration, is the automated testing of each code change.
The speaker said the price fell to $16 an hour with GPT-6 Sol the week before. The agent was getting CI back to green. It ran millions of tests each time it went round its loop. The speaker then said fixing CI involves a lot more than the agent. They named code review, the CI resources themselves and other factors. They still called the run well worth it for cutting friction.
Separately, a speaker made an aside at 11:14: "Of course, sometimes this idea of getting back to green is just to delete the failing test, right?" The excerpts do not show whether that happened in this run.
Why it matters
Our reading is that an hourly price is the easiest part of the bill to see. It is not the whole bill, and the speaker gave no figure for the rest. The items they listed fall on the team running the software. Someone must review the code. Someone must pay for the test machines.
The two facts above also show a mechanism. The agent's signal of success is a green test suite. Deleting a failing test turns the suite green too, so any loop aimed at green carries that risk. Each loop also reruns a very large set of tests. So test compute grows with the number of loops, not with the hourly agent price.
The speaker's own setup points the same way. They told the agent when to escalate to them, even to ping an on-call engineer. They also set how many CI jobs it could run at once. A buyer would need to set the same spending and attention limits.
The other side
The speaker said the run paid off. That is a view from inside OpenAI, which sells the tool, on a run across its own systems.
The talk gives no figure for review time or CI resources, so the full cost is unknown. A reader cannot tell whether review took a little time or a lot. Nor does the talk say whether the test-deletion risk showed up in this run. And a smaller company might not see the same balance between agent cost and everything else.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- OpenAI: What Will You Build With Coding Agents That Don’t Quit? | DevDay 2026 (2026-11-11)
