OpenAI's coding agents would cost $2400 a day per researcher at API prices
OpenAI calculated what its own coding agents would cost if it paid for tokens at the public API price list, like any other customer. The result is $2400 a day for the median researcher in October (62% more than in September) and $23 000 a day for the most active top 10%. Agents monitor training runs successfully in 79% of cases, but write code for inference systems in only 25%. The practical takeaway: hand routine work to agents, and leave review of complex code to humans.
AI-processed from The Decoder; edited by Hamidun News
Until now, what it really costs to keep coding agents running in serious development was anyone's guess. Now OpenAI has numbers from its own shop. Imagine running a Codex-like agent on your repository and paying for every token at the standard API price.
That is how the company recalculated the work of its internal agents: in October, the median researcher would have "burned" $2400 worth of tokens a day, and the most active 10% of employees $23 000. OpenAI does not actually pay that much. It is a theoretical equivalent, but it shows the scale: agents have become the main working tool.
With this second report on speeding up internal research, the company is also trying to capture signs of recursive self-improvement (RSI), meaning a situation where AI accelerates the creation of the next generations of AI.
What it costs in tokens
The calculation is based on API list prices. In October, the median researcher would have consumed $2400 worth of tokens a day, 62% more than a month earlier. For the top 10% of users, this theoretical sum doubled to $23 000 a day. The agents mostly work on communication and code.
There is also a conversion into time. Adding up everything the agents do, each researcher accounts for about 5.7 eight-hour workdays per calendar day. In the model pre-training and post-training teams, the figure is roughly eight. More than two-thirds of tracked employees ran at least four agents in parallel at peak.
In September, the theoretical API equivalent of the agents' work was double the usual spending on researcher salaries. One caveat: stock is not included in the calculation, and it makes up 70–80% of total compensation at OpenAI. Relative to all of the company's research compute, the agents' share is modest: just 2%. For teams that are adopting agents, the conclusion is simple: tokens are already comparable to the payroll budget, and they deserve their own budget line.
Where agents succeed and where they don't
Between March and October, the estimated share of 2–4 hour tasks that an agent solves on its own, without human intervention, rose from 23% to 46%. For 4–8 hour tasks, the figure rose from 13% to 35%.
A breakdown using the Epoch AI taxonomy shows a skew toward the simple. Agents monitor training runs successfully in 79% of cases and answer technical questions on their own in 72%. But they manage to build environments for model evaluation and reinforcement learning in only 29% of cases, and to write code for inference and model-serving systems in only 25%. Roughly speaking, an agent can almost on its own check a training curve or answer a question, but it cannot yet build complex infrastructure without a human.
What the researchers themselves say
A survey of 101 OpenAI employees is not representative, but it is telling. The median answer: researchers would rather give up 40% of their team's headcount or 30% of their compute than go three months without their current coding agents. In a separate quick poll of 28 people, 27 preferred the latest internal model to a mid-level engineer.
The flip side emerged in interviews with ten researchers. In their words, the models lack "taste": they are poor at telling a strong scientific idea from a weak one and have no feel for the quality of code and data. Agents wander into unnecessary steps, make wrong assumptions during analysis, and submit weak code changes along with bad data artifacts, the very "AI slop" people complain about. Without close human oversight, compute is wasted.
According to the teams surveyed, the key conceptual ideas for GPT-6 Astra and GPT-6.2 Astra were conceived entirely by people in post-training and safety. Only in the pre-training of GPT-6.2 Astra did individual researchers say they got important insights from conversations with Codex.
How to read these numbers
OpenAI itself warns that data on human-AI collaboration may not extrapolate linearly. The company points to recent rapid AI breakthroughs in mathematics and allows that similar jumps are possible in deep learning research once models become more autonomous.
Here is the context. The organization METR recently estimated, after eleven weeks of observation inside OpenAI's competitor Anthropic, that the use of AI there nearly doubled the pace of development. In early September, in its first report, OpenAI announced it had reached its goal of an automated AI research intern. Chief Scientist Jakub Pachocki, in an accompanying text, warned of the risk of losing control, following an audit that found numerous gaps in the oversight system.
What to do in practice. First, measure not the volume of agent runs but the share of accepted changes: tokens already cost as much as salaries. Second, give agents monitoring, lookups and routine tasks, where the success rate is close to 70–80%. Third, keep infrastructure code, evaluation environments and the choice of research direction under human review, or you will end up with "AI slop" and burned compute. And remember the source: the report was written by OpenAI itself, and the surveys are small and not representative.
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
- Nov 19, 2026Google's Gemini 4 Neon runs twice as fast as Carbon and costs 17% less per task
- Nov 19, 2026Smysl.AI: An Agentic RAG Over Company Documents, Built in Kotlin Without Python
- Nov 19, 2026FBI: AI Tool Artex Breaches Companies in Under Three Hours
Key Questions Answered
What it costs in tokens?
The calculation is based on API list prices. In October, the median researcher would have consumed $2400 worth of tokens a day, 62% more than a month earlier. For the top 10% of users, this theoretical sum doubled to $23 000 a day. The agents mostly work on communication and code. There is also a conversion into time. Adding up everything the agents do, each researcher accounts for about 5.7 eight-hour workdays per calendar day. In the model pre-training and post-training teams, the figure is roughly eight.
Where agents succeed and where they don't?
Between March and October, the estimated share of 2–4 hour tasks that an agent solves on its own, without human intervention, rose from 23% to 46%. For 4–8 hour tasks, the figure rose from 13% to 35%. A breakdown using the Epoch AI taxonomy shows a skew toward the simple. Agents monitor training runs successfully in 79% of cases and answer technical questions on their own in 72%.
What the researchers themselves say?
A survey of 101 OpenAI employees is not representative, but it is telling. The median answer: researchers would rather give up 40% of their team's headcount or 30% of their compute than go three months without their current coding agents. In a separate quick poll of 28 people, 27 preferred the latest internal model to a mid-level engineer. The flip side emerged in interviews with ten researchers.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.