Research At AI’s Jagged Frontier: Promise And Limits
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Research At AI’s Jagged Frontier: Promise And Limits on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

OpenAI’s latest internal measurements show a sharp increase in coding-agent use across its research organization since the company’s September report. The figures track runtime, spending and concurrent agents, but OpenAI says they are partial indicators—not proof that AI has accelerated research or a declaration that it plans full recursive self-improvement.

OpenAI has published updated measurements showing that researchers in its organization used coding agents for an average of 5.7 eight-hour workdays per researcher per calendar day, up from 0.5 in the company’s September analysis. The figures offer a partial view of how AI tools are being used to develop AI systems, but do not establish that research is progressing faster or that OpenAI intends to pursue full recursive self-improvement.

The report estimates that the median researcher used agentic AI tools worth about $2,400 a day at API list prices. For researchers at the 90th percentile, the estimate was about $23,000 a day. OpenAI compared these with $1,500 and $11,000, respectively, in September, describing the changes as increases of 62% and 105%. The company says these spending figures reflect several factors, including activity levels, model selection and changes to models and agent software; they are not a direct measure of research output.

OpenAI defines an “agent workday” as eight hours of agent runtime. On pretraining and posttraining teams, average agent runtime was about eight workdays per researcher per calendar day; on scaling and infrastructure teams, it was about four. By mid-October, more than two-thirds of researchers covered by the analysis had a daily peak of at least four agents running concurrently, and more than 10% averaged a peak of at least 16.

The company says the runtime measure generally captures interactive agent use and may miss some automated activity. It also reports that internal AI deployment accounts for 2% of total research compute. OpenAI compares API-equivalent agent use with non-equity researcher compensation, but that compensation measure excludes equity, which the company says generally represents 70% to 80% of researchers’ target compensation.

At a glance
reportWhen: Update published after the September re…
The developmentOpenAI published updated internal measurements showing increased coding-agent use among its AI researchers, including average runtime of 5.7 eight-hour agent workdays per researcher per calendar day.
AI Research · Briefing

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.

5.7
agent workdays per researcher per day, up from 0.5
$2,400
median daily agent use at API list prices
$23,000
daily agent use at the 90th percentile
2%
of total research compute from internal AI deployment

Agent use is rising

Eight hours of agent use counts as one agent workday.

September
  • 0.5 average agent workdays per researcher per day
  • $1,500 median daily agent use
  • $11,000 daily use at the 90th percentile
Latest report
  • 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.

Pretraining and posttrainingAbout 8
Scaling and infrastructureAbout 4

More agents work concurrently

By mid-October, many researchers had several agents active at their daily peak.

Concurrency

More than two-thirds

Had a daily peak of at least four agents working at once.

High peaks

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.

1
Activity is not output

Runtime and spending do not show whether agents produced useful work or improved research pace, quality, or success.

2
Costs are partial

API list prices reflect changes in usage, model prices, and agent behavior; the compensation comparison excludes equity.

3
Coverage is incomplete

The workday measure may miss automated activity, and the report describes its account as partial.

ContextOpenAI says the report is not an indication that it intends to pursue full recursive self-improvement.
The takeaway
Coding agents are becoming a substantial part of researchers’ workflows, but comparable measures of research outcomes are needed to show what that use achieves.

What Rising Agent Use Shows

The figures matter because they provide a rare, quantitative account of how one major AI developer is allocating agent tools inside its research organization. If coding agents can handle more tasks or run for longer, they could affect how researchers test ideas and build systems. OpenAI’s measurements show that agent activity and estimated spending have risen; they do not show how much useful work agents completed or whether the increase changed the pace or results of research.

The report also connects internal tool use to broader questions about AI-assisted research acceleration. OpenAI argues that tracking such indicators could help inform public discussion of capabilities, risks and safeguards. That is a rationale for measurement, rather than evidence that recursive self-improvement has occurred. Runtime and spending describe inputs; assessing effects would require evidence about researchers’ work, resulting systems and the contribution of other resources such as data, compute and human labor.

Products Worth Considering

Samsung Odyssey OLED G8 34-Inch Curved Ultra WQHD Gaming Monitor

Samsung Odyssey OLED G8 34-Inch Curved Ultra WQHD Gaming Monitor

Bose QuietComfort Wireless Bluetooth Noise-Cancelling Headphones, Ice Blue

Bose QuietComfort Wireless Bluetooth Noise-Cancelling Headphones, Ice Blue

As an affiliate, we earn on qualifying purchases.

From September Measures to October

The new report follows an OpenAI analysis published in September on research acceleration. The company says this update adds quantitative and qualitative measures of AI’s contribution across research areas and of the continuing role of human researchers. Its framework considers research inputs, the process of turning those inputs into model capabilities, and changes in those capabilities. The new figures focus mainly on coding agents as one research input, not on the full development process.

OpenAI says its spending comparison uses monthly API-equivalent agent use against monthly non-equity compensation. The labor-cost measure includes salary, overtime, bonuses, benefits and employer taxes, while excluding equity. The company presents the work as part of a broader effort to develop shared ways to measure AI research and development acceleration, and says other organizations have also contributed to that effort.

What the Usage Figures Cannot Prove

The report does not establish how much agent use changed research speed, quality or success rates. It does not show whether agents produced useful work, displaced researcher effort or contributed to improvements in model capabilities. Although the report describes planned analysis of researchers’ contributions and time allocation, the supplied material does not give the results of that analysis.

The measures also have limits of scope. Interactive-use estimates may miss automated activity, while API list-price valuations can change with usage, model prices and agent behavior. The compensation comparison excludes equity, and OpenAI characterizes the overall account as partial. The available information does not provide enough detail to determine how representative the researchers included are or to assess fully how the figures were calculated across teams.

Further Measures of Research Impact

OpenAI says the report is intended to contribute to shared measurement practices for AI research and development. The next useful evidence would connect tool-use indicators to researcher time allocation, completed work and changes in model capabilities, while accounting for compute, data and other inputs. The company describes further analysis of researchers’ contributions and time allocation, but the supplied material does not specify when results will be published.

For now, the update provides a snapshot of growing coding-agent use within OpenAI’s research teams. Whether that activity translates into faster or more successful research remains an open question; the published figures alone cannot answer it.

Key Questions

What did OpenAI report?

OpenAI reported that average coding-agent runtime reached 5.7 eight-hour agent workdays per researcher per calendar day, up from 0.5 in its September analysis.

Does the report show that AI research is moving faster?

No. The figures measure indicators such as agent runtime and estimated spending. They do not establish a change in research speed, quality or success rates.

What does OpenAI mean by an agent workday?

The company counts eight hours of agent use as one workday. It says this measure generally covers interactive use and may omit some automated activity.

Is OpenAI saying it plans full recursive self-improvement?

No. OpenAI says the measurements should not be read as an intention to pursue full recursive self-improvement. The report presents agent use as a topic for measurement and discussion, not proof of such a plan.

Primary source: OpenAI · via ThorstenMeyerAI.com

You May Also Like

10 Best Studio Condenser Microphones For Home-Based Recording In 2027

A buyer’s guide to 10 condenser microphones for home recording, comparing USB, XLR, controls and bundled accessories.

Rymvard Adds SAP HANA Memory To The Data Center Capacity Equation

Rymvard’s early-access 0.23 release tracks SAP HANA memory and replication capacity. Its reported tests used a simulated, not customer, estate.

What Skills-Based Matching Means For Small-Business Sales

A proposed marketplace workflow would match business buyers to listings by operating skills, then test whether better fit produces more serious offers.

How Video Games Inspire Great UX (2019)

Interest in how video game design principles inform mainstream UX work is rising among designers, though the trigger for the spike is unconfirmed.