OpenAI says its coding agents can now independently complete 46% of research tasks that would typically require a human researcher between two and four hours. The figure, based on the company’s internal measurements, represents a significant increase from 23% in March 2026 and highlights how AI agents are becoming more capable of handling complex software engineering and research-support work.

In a report published on November 18, 2026, titled Research Acceleration: The Jagged Frontier, OpenAI also reported that agents completed 35% of tasks estimated to require four to eight hours of human work, up from 13% in March. However, the company stressed that progress remains uneven. Agents perform particularly well in monitoring and technical support tasks but still struggle with more demanding engineering work, including building inference and serving systems. Human researchers continue to make key decisions and oversee the research process.

Key Takeaways

  • 46% success rate: Coding agents completed an estimated 46% of two-to-four-hour research tasks without human intervention in October.
  • Doubled since March: The success rate increased from 23% in March 2026.
  • Longer tasks: Agents completed 35% of four-to-eight-hour tasks, compared with 13% in March.
  • Strongest performance: Agents achieved a 79% success rate in monitoring AI training runs.
  • Weakest performance: Success was 25% for writing code for inference and serving systems.
  • Rising usage: Median daily agent usage per researcher reached an estimated $2,400 at API list prices in October.
  • Human oversight remains essential: OpenAI says researchers still determine priorities, evaluate results and make important decisions.

OpenAI Reports a Sharp Improvement in Coding Agent Capabilities

OpenAI’s latest findings offer an inside look at how AI coding agents are changing the company’s own research operations.

The report measures how often agents successfully complete tasks delegated by researchers without requiring human intervention. Tasks are grouped according to how long a human would typically need to complete them.

The results show that AI agents have become more effective at handling work that previously required several hours of engineering or research support.

Estimated human task durationMarch 2026October 2026
2–4 hours23%46%
4–8 hours13%35%

The improvement is notable across both task categories. The success rate for two-to-four-hour tasks doubled, while the rate for four-to-eight-hour tasks increased by 22 percentage points.

However, these figures do not mean agents can independently complete 46% of all software development projects. They describe a specific set of tasks performed within OpenAI’s research organization, using the company’s internal measurement methodology.

AI Agents Perform Better at Some Research Tasks Than Others

One of the most important findings in OpenAI’s report is that coding agent capabilities are not improving equally across every type of work.

The company describes this uneven pattern as a “jagged frontier.” An AI agent may be highly effective at monitoring a training experiment but struggle with writing complex production infrastructure code.

OpenAI’s task-level breakdown illustrates the difference.

Research taskAgent success rate
Monitoring training runs for issues79%
Answering technical questions72%
Launching training runs35%
Building evaluation, reinforcement-learning and agent environments29%
Writing code for inference and serving stacks25%

The figures cover the August 1 to October 18 measurement period.

Monitoring AI Training Runs Is a Strong Use Case

Monitoring training runs was the strongest category, with a reported success rate of 79%.

AI model training involves large computing workloads that run for extended periods. Researchers need to monitor these experiments for errors, unexpected behaviour and infrastructure problems.

Coding agents can help by examining logs, identifying potential issues and providing technical explanations.

This type of work is relatively well suited to agents because it often involves structured information, identifiable problems and established troubleshooting procedures.

Complex Infrastructure Work Remains Difficult

The success rate was considerably lower for writing code for inference and serving systems, at 25%.

Inference infrastructure is responsible for running trained AI models and delivering their outputs to users. It requires careful engineering around performance, reliability, computing resources and system scalability.

Errors can affect large numbers of users or disrupt important services.

The lower success rate suggests that agents still struggle with some of the complex engineering tasks required to operate advanced AI systems.

This difference explains why AI coding tools can be highly productive in certain workflows without being reliable replacements for experienced engineers.

OpenAI Researchers Are Increasingly Using Multiple AI Agents

The report also highlights how researchers are changing their daily working habits.

Rather than using one AI assistant for occasional coding questions, researchers are increasingly delegating multiple tasks to agents and running several sessions simultaneously.

OpenAI reported that agent workdays per researcher per day increased from 0.5 to 5.7.

The company defines one agent-workday as eight hours of agent usage. This is a measure of accumulated agent activity, not necessarily eight hours of uninterrupted work or output equivalent to a human employee.

By mid-October, more than two-thirds of researchers included in the analysis averaged a daily peak of at least four concurrent agents. More than 10% averaged at least 16.

This indicates a shift toward workflows in which researchers coordinate multiple AI agents rather than relying on a single conversational assistant.

AI Agents Are Becoming Part of the Research Workflow

OpenAI’s findings suggest that agents are increasingly involved in several stages of AI development, including:

  • Writing and modifying research infrastructure code.
  • Troubleshooting technical problems.
  • Monitoring training experiments.
  • Answering technical questions.
  • Helping researchers build evaluation environments.
  • Supporting the development and debugging of software systems.

However, agents are still more commonly used for execution-related work than for deciding what research should be pursued.

Researchers continue to determine which ideas are worth investigating, which experiments should be conducted and whether results justify further investment.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.