AIToday
AI Coding AssistantsAI Business & IndustryTHE DECODERPublished: Nov 20, 2026, 06:01 JST

OpenAI researchers value Codex over 40% of staff

Send on LINE
OpenAI researchers value Codex over 40% of staff

3 Key Points

  1. What happened

    OpenAI's report says median internal token use hit $2,400 per researcher per day in October on Codex and Dots, up 62 percent in a month; in a 101-person survey, the median would give up 40 percent of staff to keep agents.

  2. Why it matters

    That trade-off signals how deeply coding agents are now woven into OpenAI's own model-development work, with researchers preferring to lose colleagues or compute rather than the tools.

WHO IT HITSOpenAI's research and engineering staff are the immediate users affected, as the report shows their daily workflows now depend on Codex and Dots; the survey results suggest retention and team planning at AI labs may increasingly hinge on keeping such tools available.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

OpenAI's second internal research-acceleration report documents how quickly coding agents have moved from novelty to daily infrastructure inside the company. Between March and October, the estimated success rate of autonomous agents on two- to four-hour tasks rose from 23 to 46 percent, and on four- to eight-hour tasks from 13 to 35 percent. Yet the same data shows the systems mainly handle simpler work: monitoring training runs and answering technical questions go well, while building evaluation environments or writing inference and serving code succeeds far less often.

Qualitative interviews with ten researchers point to a persistent gap in what the report calls research taste. The models get lost in unnecessary steps, make wrong assumptions in analyses, and submit subpar code changes along with unusable data artifacts, making close human oversight essential to avoid wasted compute. The teams said the key conceptual ideas for GPT-6 Astra and GPT-6.2 Astra came exclusively from humans in post-training and safety, with only pretraining for GPT-6.2 Astra yielding important insights through dialogue with Codex.

OpenAI cautions that the data so far may not extrapolate linearly, pointing to recent abrupt breakthroughs in mathematics as a possibility for deep-learning research itself as models gain autonomy. The report follows METR's eleven-week look inside Anthropic, which estimated AI use there had nearly doubled the pace of development, and OpenAI's own first report in early September declaring it had reached its self-imposed goal of an automated AI research intern.

FAQ
How much are OpenAI researchers using coding agents?
At regular API list prices, the median researcher would have consumed $2,400 worth of tokens per day in October, and the top 10 percent doubled that to $23,000 per day.
What did the internal survey of OpenAI employees find?
In a non-representative survey of 101 employees, the median researcher said they would rather give up 40 percent of their team's staff or 30 percent of compute capacity than go three months without current coding agents.
Where do the agents still fall short?
Agents complete training-run monitoring 79 percent of the time and answer technical questions independently 72 percent of the time, but succeed only 29 percent on building evaluation and reinforcement-learning environments and 25 percent on writing inference and serving code.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAWS Marketplace lets AI agents buy software on a set budget