OpenAI’s internal use of artificial intelligence coding agents is growing rapidly, with researchers increasingly delegating software development, technical troubleshooting and research tasks to AI systems. The company’s latest research acceleration report reveals that the median researcher generated approximately $2,400 worth of AI agent usage per day in October 2026, calculated using standard API prices. Among the top 10% of users, the theoretical value reached approximately $23,000 per day.

The figures highlight how deeply AI agents are becoming embedded in the development of advanced AI models. OpenAI also found that its research teams now generate the equivalent of 5.7 eight-hour agent workdays per researcher per calendar day. In some pretraining and post-training teams, that figure reaches approximately eight agent workdays. The growing adoption suggests that AI systems are increasingly participating in the research and engineering processes used to build future AI models.

OpenAI’s Internal AI Agent Usage Surges

In a research update published on November 18, 2026, OpenAI shared new measurements showing how its researchers use coding agents and other AI tools.

The report examines the growing role of agentic systems in the company’s research organization. These systems can write code, investigate technical problems, monitor experiments and perform multi-step engineering tasks with varying levels of human supervision.

OpenAI found that the median researcher’s daily AI usage reached approximately $2,400 in October when calculated at standard API list prices. This represented a 62% increase from September, when the equivalent figure was approximately $1,500.

The increase was even more pronounced among the heaviest users. Researchers at the 90th percentile generated approximately $23,000 worth of daily AI usage, up 105% from around $11,000 in September.

Key Figures From OpenAI’s Research

MetricLatest Finding
Median daily AI usage value$2,400
90th-percentile daily usage value$23,000
Median usage growth from September62%
90th-percentile usage growth105%
Agent workdays per researcher per day5.7
Agent workdays in pretraining and post-training teamsApproximately 8
Researchers with peak usage of at least four concurrent agentsMore than two-thirds

These figures represent theoretical API-equivalent costs rather than actual internal expenditure. OpenAI operates its own infrastructure and uses internal deployments, so the list-price calculation should not be interpreted as the amount the company pays for each researcher’s AI usage.

AI Agents Now Perform More Work Than a Conventional Eight-Hour Day

One of the most interesting measurements in the report is the number of agent workdays generated by researchers.

OpenAI defines one agent workday as eight hours of AI agent runtime.

According to the company, agent workdays increased from approximately 0.5 per researcher per day to 5.7, representing an elevenfold increase.

In pretraining and post-training teams, average agent runtime reached roughly eight workdays per researcher per calendar day.

This does not mean an individual researcher works eight additional human shifts. Multiple AI agents can operate simultaneously, allowing a researcher to delegate several activities while continuing to perform other work.

For example, a researcher could ask one agent to investigate a software bug, another to write evaluation code and a third to analyze experimental results.

The researcher can then review the outputs and decide which changes should be incorporated into the project.

Researchers Are Running Multiple AI Agents Simultaneously

OpenAI’s report also shows that researchers are increasingly using several agents at the same time.

By mid-October, more than two-thirds of the researchers included in the analysis averaged a daily peak of at least four concurrent agents.

More than 10% averaged at least 16 concurrent agents.

This development reflects a shift from using AI as an assistant that responds to individual prompts toward treating it as a team of specialized digital workers.

Concurrent agents can divide a larger assignment into smaller tasks. One system might inspect a codebase, another might run tests, and another might prepare documentation.

However, managing multiple agents also creates new challenges. Their outputs must be coordinated, checked for conflicts and reviewed before being incorporated into important research projects.

AI Agents Are Becoming Better at Completing Research Tasks

The growing usage is accompanied by measurable improvements in autonomous task completion.

OpenAI evaluated how often AI agents could independently complete tasks that would typically require several hours of human work.

Between March and October 2026, the estimated success rate for tasks requiring two to four hours of human effort increased from 23% to 46%.

For tasks requiring four to eight hours, the success rate rose from 13% to 35%.

Task DurationMarch 2026October 2026
2–4 hours of human work23%46%
4–8 hours of human work13%35%

These results indicate that AI agents are becoming more capable of handling longer assignments without continuous human intervention.

However, their performance varies significantly depending on the type of work.

Where AI Agents Perform Best

OpenAI’s task-level analysis found that agents achieved particularly strong results in routine research-support activities.

Research ActivityAutonomous Agent Success Rate
Monitoring training runs79%
Answering technical questions72%
Launching training runs35%
Building evaluation and reinforcement-learning environments29%
Writing inference and serving-system code25%

The figures reveal an important limitation: agents are considerably more reliable at monitoring and answering questions than at building complex infrastructure.

Writing inference and serving code, for example, requires a detailed understanding of system architecture, performance constraints and interactions between software components.

This is why OpenAI continues to emphasize the importance of human researchers.

AI Agent Costs Have Exceeded Researchers’ Non-Equity Compensation

Perhaps the most striking finding concerns the relationship between AI usage and human labor costs.

OpenAI compared the theoretical API-equivalent value of agent usage with the non-equity compensation paid to researchers.

As of September 2026, the estimated value of agentic AI usage in the research organization was approximately twice its non-equity compensation.

However, this comparison requires an important qualification.

OpenAI’s researchers receive equity compensation, which the company says generally accounts for 70% to 80% of their target compensation. The comparison excludes that equity component.

It also does not mean OpenAI is spending twice as much cash on AI agents as it spends on researchers.

The calculation is designed to show how much AI capacity the organization is consuming relative to one component of its human labor costs.

Despite the high theoretical value, internal AI deployment through Codex, ChatGPT and API usage represented only 2% of OpenAI’s total research computing resources.

Researchers Still Need Human Judgment

Despite the rapid increase in AI agent usage, OpenAI’s report does not suggest that researchers can be replaced entirely.

AI agents can write code, conduct experiments and troubleshoot technical problems, but they still struggle with activities requiring deeper engineering judgment.

Qualitative interviews with researchers also highlighted problems with low-quality AI-generated output, often described as “AI slop.”

This can include code that appears functional but contains architectural weaknesses, unnecessary complexity or errors that become visible only during more demanding testing.

Researchers must therefore review generated work, identify promising ideas and determine which experiments are worth pursuing.

Another important limitation is research judgment, sometimes described as research taste. This involves recognizing which problems are important, selecting promising approaches and deciding whether an experimental result represents a meaningful improvement.

These decisions remain difficult to automate reliably.

OpenAI Is Exploring Recursive Self-Improvement

The growing use of coding agents is also connected to a larger ambition in AI research: recursive self-improvement.

The concept describes a process in which AI systems help accelerate the development of more capable future AI systems.

For example, an AI agent could help researchers write software, develop evaluation tools, identify training problems and improve infrastructure. Those improvements could then help researchers build stronger models.

Stronger models could subsequently become more effective at assisting with the development of the next generation.

This creates the possibility of a feedback loop in AI research.

However, OpenAI’s latest report does not demonstrate that fully autonomous recursive self-improvement has already been achieved. Human researchers, computing resources and experimental validation remain important constraints.

The Bigger Picture

OpenAI’s internal data offers a glimpse into how the economics of AI research are changing.

The traditional research process depends heavily on human engineers and scientists writing code, managing experiments and investigating technical problems. AI agents are increasingly taking over parts of these workflows, allowing researchers to run more activities in parallel.

The theoretical value of AI usage exceeding non-equity researcher compensation is an indication of how much computational work is being delegated to software systems.

However, the gap between strong performance on routine tasks and weaker performance on complex engineering demonstrates why human oversight remains essential.

The most important development may not be the replacement of researchers but the expansion of what a relatively small research team can accomplish with multiple AI agents.

Looking Ahead

The next stage of AI development will depend partly on whether agents can move beyond routine coding and monitoring tasks toward more complex research activities. Improvements in long-term planning, software engineering, evaluation and independent problem-solving could allow AI systems to take responsibility for larger portions of the model-development process.

For OpenAI and other AI developers, the strategic opportunity is to build systems that accelerate research without sacrificing reliability or human control. If AI agents continue improving, they could become an increasingly important part of the infrastructure used to develop future generations of artificial intelligence.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.