OpenAI Internal Report Says Agents’ API-Equivalent Value Has Reached Twice Researchers’ Non-Equity Compensation
OpenAI’s internal report, “Research Acceleration: The Jagged Frontier,” shows that Agents can now independently complete 46% of research tasks that previously took humans two to four hours, up from 23% in March this year. The API list-price equivalent value of Agent usage by the median researcher has risen from $1,500 to $2,400 per day, with the overall API-equivalent value reaching twice researchers’ non-equity compensation. Agents perform well on communication and analysis tasks, but their completion rate for building tasks is 43%, while their success rate for writing inference and service code is only 25%.
Why read this
SelectedAI score 79
The report quantifies how much Agents accelerate research tasks while also revealing the limits of their capabilities in system building and research judgment.
Topics
Body · AI translation
Is AI More Expensive Than Humans? OpenAI: Internal Agent Costs Have Reached Twice Researchers’ Compensation. Download the QbitAI App for an Exclusive First Look. November 20, 2026, 12:58:26, from Beijing. Updated November 20, 2026, 13:27:20
By Wen Le, from Aofeisi
QbitAI | Official WeChat Account: QbitAI
To what extent are OpenAI researchers using AI now?
The Agents called by each person every day are worth an astonishing $2,400 at API list prices?!
The top 10% of researchers by usage even consume as much as $23,000 in a single day—about RMB 160,000.
Wow... These researchers really do not hold back on Tokens when using AI.
OpenAI has released its latest internal Agent research report, “Research acceleration: The jagged frontier,” offering an exhaustive look at how its own researchers use AI.
First, here is a one-sentence summary:
OpenAI’s internal Agents can now independently complete nearly half—46%—of research tasks that previously took humans two to four hours.
In March this year, that figure was only 23%; in just seven months, it doubled.
For research work that originally took humans four to eight hours, the rate at which Agents could complete it autonomously rose from 13% to 35%.
OpenAI asked 101 researchers:
If you were denied access to the most advanced Agents for the next three months, what would you be willing to give up to regain access?
Respondents were generally willing to trade 40% of their team’s headcount or 30% of their computing resources.
The median researcher’s daily spending on Agents rose from $1,500 to $2,400.
For the top 10% of users, it soared from $11,000 to $23,000.
The report also specifically calculated a metric called an “Agent workday.”
The formula is simple:
Every eight cumulative hours an Agent runs counts as one workday.
Converted into workload, each researcher can now drive an average of 5.7 Agent workdays per day.
That is like having nearly six tireless AI underlings behind every researcher, standing by 24/7.
This figure is a full 11 times the previous benchmark.
And everyone has already grown accustomed to working in “swarm mode.”
More than two-thirds of researchers run at least four Agents simultaneously every day; more than 10% are even heavier users, running at least 16 at once.
And that is not all. One comparison in the report is particularly striking:
The API-equivalent value of Agents has now reached twice researchers’ non-equity compensation.
(Although that compensation excludes the major component of equity, it is still alarming...)
In other words, within OpenAI, the value of the labor provided by AI has already surpassed that of humans on paper.
The researchers themselves agree.
An internal survey posed a brutal question:
If you had to go three months without the most advanced coding Agent, what would you be willing to sacrifice to get it back?
On one side is an AI Agent that can work for you 24/7; on the other are 40% of the living, breathing colleagues on your team—
Colleagues and compute, you are free now (doge).
This is the central question, but the answer is: The jagged frontier—uneven and inconsistent.
Agents perform well on tasks such as “communication” (69%) and “analysis” (58%), but lag considerably on the most crucial category, “building” (43%).
Put simply, they are good at researching information, processing data, writing emails, and monitoring experiments.
But ask them to build a complex system from scratch, and they can easily become lost.
The data shows a 79% success rate for monitoring training jobs, but only 25% for writing inference and serving code.
This exposes a key problem—one researchers repeatedly complained about in interviews:
AI Agents lack “research taste.”
Models generally cannot independently formulate a well-defined problem that they themselves cannot solve.
“I’ve tried telling it, ‘Do this task just like that task,’ and it has backfired on me many times, because it never gets it quite right and I have to fix it myself.”
The researchers’ feedback was consistent:
Agents are excellent tools for implementing code and debugging, but they cannot formulate good questions themselves, distinguish good code from bad, or judge the quality of data.
Human researchers are responsible for generating ideas, setting direction, defining problems, making key decisions, and performing final quality control.
AI Agents, meanwhile, are like supercharged “hands” and “eyes,” responsible for rapidly implementing ideas and carrying out huge volumes of tedious testing and analysis.
As one researcher put it, evaluating whether an open-source project was useful used to take at least a week or two.
Now? It can be up and running with a conclusion in one day, or a few days at most.
Of course, this boundary is also becoming blurred.
The report notes that during the development of GPT-6.2 Astra, the pretraining team attributed the generation of a key idea to collaborative conversations with Codex for the first time.
Although the idea was the researcher’s, Codex proposed a crucial modification that made the entire approach work.
Code output: In 2026, the number of lines of code produced by each active contributor continued to surge compared with the period before 2025.
Experimental efficiency: The number of experiments run by each active experimenter reached 1.68 times the 2025 benchmark.
More direct evidence comes from an internal OpenAI capability index (OECI).
Since August 2025, the frontier of OpenAI model capabilities has advanced by 16.4 points per year, exceeding the previous industry trend of 14 points per year.
But OpenAI’s report remained measured in its conclusion.
The focus is Safety and Alignment.
The increase in workload in this field was the highest of any area.
Large numbers of researchers, both newly hired and internally transferred, are pouring into safety projects.
“The jagged frontier” may describe not only the unevenness of technological capabilities, but also the reality of our era: stepping on the accelerator while searching for the brakes.
“Special statement: The above content (including any videos, images, or audio) was uploaded and published by users of Dafeng Hao, a self-media platform operated by Phoenix New Media. This platform merely provides information storage services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services.”
AI Trends