🔍 Read the full analysis: AI’s Jagged Frontier: Rethinking How Research Moves on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
OpenAI’s latest internal report says average coding-agent runtime in its research organization rose from 0.5 to 5.7 eight-hour agent workdays per researcher per calendar day since its September report. The company calls the figures partial measures of tool use, not evidence of faster research or an intention to pursue full recursive self-improvement.
OpenAI has published new internal measurements showing that coding agents ran for an average of 5.7 eight-hour workdays per researcher per calendar day in its research organization, up from 0.5 in its September analysis. The report documents increased use of AI tools in developing AI systems, but OpenAI says the figures are partial indicators of activity and do not establish that research is progressing faster.
OpenAI’s report measures agent use through runtime, estimated spending and the number of agents researchers run concurrently. It defines an “agent workday” as eight hours of agent use, so the average of 5.7 is a measure of total runtime per researcher across a calendar day, not a claim that an individual agent works a conventional 5.7-day shift. The measure generally covers interactive use and may miss some automated activity, the company says.
Estimated daily agent spending, valued at API list prices, was about $2,400 for the median researcher and about $23,000 for researchers at the 90th percentile. OpenAI compared those amounts with $1,500 and $11,000, respectively, in September, describing the increases as 62% and 105%. The company says spending changes reflect several factors, including activity, model choice, and changes to models and agent software; they are not direct measures of work produced.
Use varied across teams: average runtime was about eight agent workdays per researcher per day on pretraining and posttraining teams, and about four on scaling and infrastructure teams. By mid-October, more than two-thirds of researchers in the analysis had a daily peak of at least four agents running concurrently; more than 10% averaged a peak of at least 16. OpenAI also reported that internal deployment represented 2% of total research compute.
Research Acceleration The Jagged Frontier
OpenAI reports heavier use of coding agents in its research organization. The figures show activity and resource use, but do not establish faster or better research.
Agent use is rising
Eight hours of agent use counts as one agent workday.
- 0.5 average agent workdays per researcher per day
- $1,500 median daily agent use
- $11,000 daily use at the 90th percentile
- 5.7 average agent workdays per researcher per day
- $2,400 median daily agent use
- $23,000 daily use at the 90th percentile
Use differs across teams
Average agent runtime varies by research area.
More agents work concurrently
By mid-October, many researchers had several agents active at their daily peak.
More than two-thirds
Had a daily peak of at least four agents working at once.
More than 10%
Averaged at least 16 agents at their daily peak.
What the figures leave open
The measures capture inputs to research, not the effects on its outcomes.
Runtime and spending do not show whether agents produced useful work or improved research pace, quality, or success.
API list prices reflect changes in usage, model prices, and agent behavior; the compensation comparison excludes equity.
The workday measure may miss automated activity, and the report describes its account as partial.
What the Usage Measures Can Show
The figures offer a rare view into how one major AI lab is allocating AI tools within its own research organization. As coding agents take on tasks or run for longer periods, they could change how researchers test ideas and build systems. The increase in runtime and spending shows that agent use has expanded by OpenAI’s reported measures; it does not, on its own, show that research has become faster, more successful or more capable.
That distinction matters to discussions about recursive self-improvement—the possibility that AI systems could help accelerate the development of later systems. OpenAI argues that tracking AI’s contribution to research can inform debates about capabilities, risks and safeguards. It explicitly says this update should not be read as an intention to pursue full recursive self-improvement. The report is a measurement effort, not a finding that such a feedback loop has been achieved.
The numbers also have limits as comparisons of resources. API list prices are an estimate of the value of agent use, not necessarily the company’s actual costs. OpenAI’s comparison with researcher compensation excludes equity, which the company says generally represents 70% to 80% of target compensation. The company describes agent use as one part of research alongside human labor, data and compute, rather than a complete account of the inputs behind new systems.
Products Worth Considering

Apple AirPods Pro

Jackery Fridge Guard 1024Wh Backup Power Station with Mount

Apple Watch Series 12 GPS Smartwatch, 42 mm
As an affiliate, we earn on qualifying purchases.
From September’s Research Baseline
The update follows an OpenAI report published in September on research acceleration. The company says the newer analysis adds quantitative and qualitative measures of AI’s contribution across research areas, as well as continued human involvement. Its framework separates research inputs, the process of turning those inputs into model capabilities, and resulting changes in capabilities. The reported agent-use figures focus mainly on coding agents as one input.
OpenAI says its spending comparison tracks monthly API-equivalent agent use against monthly non-equity compensation. The compensation measure includes salary, overtime, bonuses, benefits and employer taxes, but excludes equity. That exclusion makes the ratio narrower than a comparison with researchers’ full target compensation. The report’s runtime measure also generally captures interactive agent use, so it may not include all automated work.
The company presents the work as part of a broader effort to develop shared measurement practices. It says other organizations have contributed to efforts to measure AI research and development acceleration. OpenAI’s stated rationale is that clearer reporting could help inform public discussion of how oversight and safeguards may need to respond as AI capabilities change.
What Runtime Cannot Establish
The reported figures do not show how much coding agents have changed the pace, quality or success rate of OpenAI’s research. Runtime and estimated spending cannot establish whether agents produced useful work, replaced or supplemented researcher effort, or contributed to changes in model capabilities. The supplied material does not include the results of the planned analysis of researchers’ contributions and time allocation.
Some details needed to interpret the measures also remain unclear from the material provided, including how representative the researchers in the analysis are and how the measures were calculated across teams. Interactive-use tracking may omit automated activity, while the spending estimates can shift with model prices and agent behavior as well as usage. The reported compensation comparison excludes equity. OpenAI describes the overall account as partial, and the available figures do not permit a full assessment of research costs or productivity.
Further Reporting on Research Inputs
OpenAI says it intends the report to contribute to shared norms for measuring AI’s role in research and development. The next useful developments would include fuller reporting on other research inputs—such as human work, data and compute—and results that connect tool use to researcher contributions or measurable changes in capabilities. The supplied material does not give a date for further analysis.
For now, the update adds a new internal snapshot to the September baseline. Readers can track future reports for changes in runtime, spending and concurrent use, while treating those indicators separately from evidence about research outcomes. Whether broader measurement efforts will produce consistent comparisons across organizations remains open.
Key Questions
What is an agent workday?
OpenAI counts eight hours of agent use as one workday. The reported 5.7 average is total agent runtime per researcher per calendar day, not a conventional work schedule for one agent or one researcher.
Does the report show that OpenAI’s research is moving faster?
No. The report measures indicators such as runtime, estimated spending and concurrent agents. OpenAI says those measures do not establish research speed, quality or success.
How much did estimated agent spending rise?
At API list prices, estimated daily use was about $2,400 for the median researcher and $23,000 for researchers at the 90th percentile. OpenAI compared those amounts with $1,500 and $11,000 in September and described the increases as 62% and 105%, respectively.
Does the report mean OpenAI plans full recursive self-improvement?
No. OpenAI says the measurements are intended to help track AI’s contribution to research and should not be read as an intention to pursue full recursive self-improvement.
Primary source: OpenAI · via ThorstenMeyerAI.com
