Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableTNW→ original

OpenAI's Agents Now Complete 46% of 2–4 Hour Research Tasks Solo, Up from 23%

OpenAI has published figures showing how its coding agents are taking on research work. For tasks that take a person two to four hours, such as digging through training logs, agents now finish the job alone in 46% of cases, up from 23% in March. On tasks lasting four to eight hours, the share rose from 13% to 35%. Mid-length routine work can now go to agents, while complex server-side code is better left to humans.

AI-processed from TNW; edited by Hamidun News
OpenAI's Agents Now Complete 46% of 2–4 Hour Research Tasks Solo, Up from 23%
Source: TNW. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTNW ↗
Editor-in-Chief
Key metric #1The median researcher spends $2,400 a day on agents at API prices
Solo success79% for training monitoring, 25% for inference code
OpenAI's caveatresearchers could change which tasks they hand to agents

OpenAI says its coding agents now complete 46% of research tasks on their own, tasks that would take a human researcher two to four hours. Examples include "go through the logs of a failed training run and find the cause" or "write a script to check a new model." In March, the share of such tasks completed successfully was 23%, so it doubled in a few months.

For larger tasks, those requiring four to eight hours of human work, the rate rose from 13% to 35%. The report came out on Wednesday and is the company's second on how AI is accelerating its own research, after one in September. The figures are useful for anyone deciding how much work can be handed to agents.

What agents already do on their own

Results depend heavily on task type, so OpenAI calls its progress "jagged." Agents succeed alone in 79% of cases when monitoring training runs and in 72% when answering technical questions. But they manage only 25% of the time on code for inference systems (that is, running and serving finished models), and 29% on building environments for evaluating and training models. Put simply, agents are good at monitoring and answering questions, and still poor at writing subtle infrastructure code.

OpenAI also includes a caveat of its own. As agents improved, researchers could change the set of tasks they entrusted to them. The rising success rate therefore reflects both model progress and new habits among people.

What it costs

Based on API pricing, the median OpenAI researcher used agents at roughly $2,400 a day in October. In September it was $1,500, a rise of 62%. Among the heaviest users, at the 90th percentile (the level only the top 10% of staff reach or exceed), daily spending reached about $23,000, compared with $11,000 a month earlier.

In time terms, agent usage per researcher grew elevenfold, from 0.5 to 5.7 "agent workdays" per day. OpenAI counts every eight hours of agent work as one workday. By mid-October, more than two-thirds of researchers were running at least four agents simultaneously at peak moments, and over 10% were running at least 16.

In September, the cost of using agents in the research division was roughly double the researchers' salary costs. Stock options are not included in that figure, and according to OpenAI they make up 70–80% of target compensation. Internal agents consumed only 2% of research compute.

Ideas still come from people

Researchers were asked where the main ideas behind GPT-6 Astra and the next version came from. Post-training and safety specialists credited people with all the ideas, saying models helped implement them. Pretraining specialists for GPT-6.2 Astra said they developed ideas in dialogue with Codex, and one noted that the model suggested key insights.

In interviews, researchers say agents still have poor taste: in ideas, code, data and writing alike. That is why they need close supervision. One pretraining specialist admitted that they currently consider their own taste better, but is not sure that will last long.

In a survey of 101 OpenAI employees, the company asked what each would give up to have three months of access to frontier coding agents. The median answer: 40% of their team's headcount or 30% of its compute. The company itself cautions that self-assessments may be inaccurate and that the sample was not gathered with rigorous methodology.

What else is in the report, and what practitioners should do

OpenAI proposes that labs be required to track and publicly report signs of recursive self-improvement, when AI accelerates the creation of the next AI. The company clarifies that it does not plan to pursue full self-improvement. To measure overall pace, OpenAI uses an internal version of the Epoch AI Capabilities Index. By that measure, the frontier has grown 16.4 points per year since August 2025, whereas Epoch saw 14 points per year from late 2024. The company notes that its models score higher on its own index than on Epoch's.

For practitioners, the takeaway is simple. Agents can already be trusted with monitoring, log analysis, answering technical questions, and tasks lasting a few hours. Critical infrastructure code is still better kept under human control. And it is worth budgeting in advance: heavy agent use runs to thousands of dollars per person per day.

Story

Key Questions Answered

What agents already do on their own?

Results depend heavily on task type, so OpenAI calls its progress "jagged." Agents succeed alone in 79% of cases when monitoring training runs and in 72% when answering technical questions. But they manage only 25% of the time on code for inference systems (that is, running and serving finished models), and 29% on building environments for evaluating and training models. Put simply, agents are good at monitoring and answering questions, and still poor at writing subtle infrastructure code.

What it costs?

Based on API pricing, the median OpenAI researcher used agents at roughly $2,400 a day in October. In September it was $1,500, a rise of 62%. Among the heaviest users, at the 90th percentile (the level only the top 10% of staff reach or exceed), daily spending reached about $23,000, compared with $11,000 a month earlier. In time terms, agent usage per researcher grew elevenfold, from 0.5 to 5.7 "agent workdays" per day. OpenAI counts every eight hours of agent work as one workday.

What are the key details on Ideas still come from people?

Researchers were asked where the main ideas behind GPT-6 Astra and the next version came from. Post-training and safety specialists credited people with all the ideas, saying models helped implement them. Pretraining specialists for GPT-6.2 Astra said they developed ideas in dialogue with Codex, and one noted that the model suggested key insights. In interviews, researchers say agents still have poor taste: in ideas, code, data and writing alike. That is why they need close supervision.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…