OpenAI says its coding agents can now independently complete 46% of research tasks estimated to require two to four hours of human work. The figure represents a significant improvement from March 2026, when agents successfully completed just 23% of comparable tasks without human intervention. The findings highlight how quickly AI agents are becoming capable of handling complex technical work inside one of the world’s leading AI research organisations.
The results were published in OpenAI’s November 18, 2026 report, Research acceleration: The jagged frontier. The study examines how coding agents are changing the company’s internal research workflows, how much researchers are spending on AI tools and which activities still require substantial human involvement. OpenAI also reported that median daily agent usage among its researchers reached an estimated $2,400 at API list prices in October, demonstrating the growing importance of autonomous AI systems in advanced research.
Key Takeaways
- OpenAI’s coding agents independently completed 46% of assigned two-to-four-hour research tasks in October, compared with 23% in March.
- Success on tasks estimated to require four to eight hours increased from 13% to 35%.
- Agents performed particularly well at monitoring AI training runs, with a reported success rate of 79%.
- Their success rate was only 25% for writing code for inference and serving systems.
- Median daily agent usage reached approximately $2,400 per researcher at API list prices.
- Researchers at the 90th percentile used approximately $23,000 worth of agentic AI services per day.
- OpenAI says human researchers remain essential for choosing research ideas, reviewing results and directing experiments.
Coding Agents Now Complete Nearly Half of Multi-Hour Research Tasks
The most important finding in OpenAI’s report is the improvement in autonomous task completion.
The company measured how often coding agents could finish assigned research activities without a human stepping in to intervene.
For tasks estimated to require two to four hours of human work, the success rate increased from 23% in March to 46% in October.
For more demanding tasks estimated to require four to eight hours, the success rate rose from 13% to 35%.
| Estimated human task duration | March 2026 | October 2026 |
|---|---|---|
| 2–4 hours | 23% | 46% |
| 4–8 hours | 13% | 35% |
These figures suggest that AI agents are becoming increasingly capable of handling assignments that previously required researchers to spend several hours writing code, investigating technical problems or conducting experiments.
However, the figures need to be interpreted carefully. OpenAI estimates task duration based on the work involved, and its researchers may have changed which assignments they delegated to agents as the technology improved.
The company therefore cautions that the increase reflects both improvements in AI capabilities and possible changes in how researchers select tasks for delegation.
AI Agents Perform Better at Some Research Tasks Than Others
OpenAI describes the current state of agent capabilities as a “jagged frontier”.
The term refers to the uneven performance of AI systems across different types of work. An agent might perform exceptionally well at one technical assignment while struggling with another task that appears similarly complex.
The company’s task-level analysis, covering August 1 to October 18, 2026, demonstrates this difference.
| Research activity | Autonomous agent success rate |
|---|---|
| Monitoring training runs for issues | 79% |
| Answering technical questions | 72% |
| Writing code for inference and serving stacks | 25% |
| Building evaluation, reinforcement-learning and agent environments | 29% |
The results suggest that agents are particularly useful for monitoring experiments, answering technical questions and analysing information.
However, they remain less reliable when required to build complex infrastructure or create the environments used to train and evaluate AI models.
This distinction matters because developing advanced AI systems requires much more than writing ordinary software.
Researchers must design experiments, build evaluation systems, manage computing infrastructure, analyse results and determine whether a proposed improvement actually works.
AI agents can accelerate individual parts of this process, but weaknesses in important technical areas can still limit overall productivity.
OpenAI Researchers Are Spending Thousands of Dollars on AI Agents Every Day
Another major finding concerns the amount of money researchers are spending on AI tools.
At API list prices, the median OpenAI researcher used approximately $2,400 worth of agentic AI services per day in October 2026.
That represented a 62% increase from September, when the median was approximately $1,500.
The most intensive users recorded even higher figures.
Researchers at the 90th percentile used approximately $23,000 worth of agentic AI services per day, compared with $11,000 in September, an increase of 105%.
| Usage measure | September 2026 | October 2026 |
|---|---|---|
| Median daily usage | $1,500 | $2,400 |
| 90th-percentile daily usage | $11,000 | $23,000 |
These are estimated values calculated using API list prices, not necessarily the actual cash expenditure incurred by OpenAI.
The increase also reflects several factors, including more frequent agent use, shifts toward more expensive models and changes in model behaviour and the software systems coordinating agents.
Nevertheless, the figures show how important AI tools have become inside OpenAI’s research organisation.
The company is increasingly using AI systems as part of the infrastructure required to develop future AI models.
Researchers Are Running Multiple AI Agents at the Same Time
OpenAI also reported a significant increase in concurrent agent usage.
Rather than asking one AI system to complete an entire research assignment, researchers increasingly operate several agents simultaneously.
One agent might investigate a technical issue, another could write or modify code, while a third analyses experimental results.
By mid-October, more than two-thirds of the researchers included in the analysis averaged a daily peak of at least four agents working concurrently.
More than 10% averaged at least 16 concurrent agents.
OpenAI also measured the amount of time agents spent working.
The company defines one agent workday as eight hours of agent activity. Using that measure, average agent workdays per researcher per day increased from 0.5 to 5.7.
That represents an approximately elevenfold increase.
Teams focused on pretraining and post-training averaged roughly eight agent workdays per researcher per calendar day. Teams working on scaling and infrastructure averaged around four.
The figures demonstrate how multiple AI agents can operate in parallel, allowing researchers to delegate different parts of a technical project simultaneously.
However, more agent activity does not automatically mean proportionally greater research output. The quality of the work, the amount of human review required and the success of experiments remain important.
OpenAI Researchers Say They Would Sacrifice Resources to Keep AI Agents
The report also includes a survey of 101 OpenAI research staff members conducted between late August and early September 2026.
Researchers were asked what they would be willing to give up to restore access to advanced coding agents if their teams lost access for three months.
The median responses were:
- 40% of team headcount: Researchers indicated they would trade this amount to retain access to frontier coding agents.
- 30% of team computing resources: This was the median amount researchers said they would sacrifice in a separate question.
- 50% of team computing resources: This was the median response when researchers considered the expected improvements in AI tools by January 2027.
The survey also asked researchers how long their teams would need to complete work equivalent to their previous month’s output without access to advanced coding agents.
The median response was three months.
These results suggest that researchers consider AI agents valuable productivity tools.
However, the survey was based on self-reported opinions and an ad-hoc sampling strategy. It should not be treated as a controlled measurement of actual productivity or a prediction that companies can directly replace 40% of their employees with AI.
Human Researchers Still Control Important Decisions
Despite the progress, OpenAI’s report does not suggest that its research organisation has become fully autonomous.
The company found that researchers continue to play a central role in selecting ideas, designing experiments and evaluating results.
During interviews, researchers described several problems with agent-generated work.
These included:
- AI agents producing code that required substantial correction.
- Models proposing research ideas that were poorly formulated.
- Agents spending too much time on unnecessary steps.
- Generated datasets containing unreliable or unusable information.
- Incorrect assumptions appearing in analysis and experimental results.
One researcher described a workflow in which they ask an agent to complete a single example first, review the result and make corrections before allowing it to perform the remaining work.
This approach reflects a practical limitation of current AI systems: they can generate large amounts of work quickly, but human judgment is still needed to ensure that the work is useful and correct.
OpenAI also found that researchers continued to originate important ideas behind its GPT-6 Astra release. For GPT-6.2 Astra, some pretraining researchers described developing ideas in collaboration with Codex.
This suggests that AI is increasingly contributing to research implementation and, in some cases, idea development, while humans remain involved in deciding what is worth pursuing.
AI Research Output Is Increasing
OpenAI also examined whether the growing use of coding agents corresponds with measurable changes in research activity.
The company tracked code changes, research experiments and model capability improvements.
For the four weeks ending October 18, 2026, the number of logged experiments per active experimenter averaged 1.68 times the 2025 baseline.
OpenAI also reported that lines of code changed per active contributor had continued to increase sharply during 2026 compared with the pre-2025 average.
However, the company cautioned that more code changes and experiments do not necessarily mean faster scientific progress.
Some experiments may not produce useful findings, and generating more code does not guarantee that the underlying AI models become more capable.
The company therefore introduced an internal metric called OpenAI ECI, based on the Epoch Capabilities Index methodology, to track changes in model capabilities.
OpenAI reported that its model capability frontier had increased by 16.4 points per year since August 2025 under this internal measure. It compared this with an Epoch-observed trend of 14 points per year since late 2024, while noting that differences in benchmark composition affect comparisons.
OpenAI Increases Its Focus on AI Safety and Alignment
The growing role of autonomous agents also creates new safety challenges.
OpenAI reported that documented researcher workweeks devoted to Safety and Alignment increased by 77% between January and September 2026.
The increase came from both new hires and existing researchers shifting their attention toward safety-related work.
The company said the increase accelerated after its response to the Hugging Face incident in July 2026.
Safety and alignment research includes work on monitoring AI behaviour, evaluating models, preventing reward hacking and ensuring that advanced systems follow intended instructions.
The findings highlight a central challenge for AI developers: as agents become more capable of carrying out complex tasks independently, researchers need stronger systems for monitoring their behaviour and identifying failures.
OpenAI also said it believes AI laboratories should track and publicly report indicators of recursive self-improvement, a process in which AI systems contribute to developing more capable future AI systems.
The company stressed that publishing its research acceleration measurements should not be interpreted as an intention to pursue full recursive self-improvement.
What the Findings Mean for Software Developers and Businesses
OpenAI’s findings have implications beyond AI research laboratories.
Coding agents are increasingly being used in software development, debugging, testing, documentation and technical analysis.
For businesses, the ability to delegate multi-hour assignments to AI systems could change how software teams organise their work.
Developers may increasingly spend less time writing routine code and more time defining requirements, reviewing generated solutions, testing software and managing complex systems.
Startups could also use AI agents to complete certain technical assignments with smaller teams, potentially reducing the time required to build early products.
However, the report does not establish that AI agents can independently replace experienced software engineers.
The relatively low success rates for inference and serving infrastructure demonstrate that important engineering tasks remain difficult for current systems.
The practical opportunity is therefore likely to involve human-AI collaboration rather than complete automation across every stage of software development.
The Bigger Picture
OpenAI’s findings offer an important look at how artificial intelligence is beginning to influence the process of developing artificial intelligence itself. Coding agents are no longer limited to generating short pieces of software or answering technical questions. They are increasingly being used to monitor experiments, build tools, analyse data and complete assignments that previously required several hours of human work.
This creates the possibility of a feedback loop in which AI systems help researchers develop more capable AI systems, which in turn become more useful research assistants. However, the current evidence shows that this process remains uneven. Human researchers still choose important research directions, evaluate results and correct mistakes. The challenge for the industry will be to improve autonomous capabilities while ensuring that human oversight and safety systems remain effective.
Looking Ahead
OpenAI plans to continue measuring the impact of coding agents on research productivity, model capabilities and safety work. Future reports could provide additional evidence about whether AI-driven research acceleration is translating into faster improvements in advanced models. The company also intends to track differences between capability development and safety research, an important issue as AI systems become more autonomous.
For businesses and software developers, the findings suggest that coding agents could become increasingly important tools for completing complex technical assignments. However, their value will depend on reliability, the quality of generated work and the availability of human review. The 46% success rate is a significant milestone, but it also demonstrates that autonomous AI agents still have considerable room to improve before they can independently handle most multi-hour research assignments.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



