SIAgentNews.com
← All news

OpenAI agents complete nearly half of multi-hour research tasks

OpenAI's internal data shows a significant jump in the ability of autonomous agents to finish complex research assignments without human intervention.

· First reported by Mezha.net

46%Independent completion of 2-4 hour research tasks

The short version

  • Agents completed 46% of research tasks requiring 2-4 hours of human work, up from 23% in March 2026.
  • Success rates for 4-8 hour tasks rose from 13% to 35% over the same period.
  • Communication and analytical tasks saw higher success rates than software development.
  • Daily spending on coding agents reached $23,000 at the 90th percentile in October 2026.
  • Human oversight remains essential as agents still struggle with research judgment and software quality.

Why it matters

This progress suggests that AI is moving from simple chat interactions to performing multi-step, multi-hour workflows independently. While humans are still needed for high-level judgment, the speed at which agents can handle routine research infrastructure is accelerating.

Rapid gains in autonomy

OpenAI recently released internal assessments showing that its autonomous agents are becoming significantly more capable of finishing complex assignments. According to reports from Mezha.net, these agents independently completed 46% of research tasks that would typically take a human between two and four hours to finish. This is a notable jump from the 23% completion rate recorded in March 2026.

The company also tracked tasks that require even more significant time investments, specifically those estimated to take four to eight hours. For these longer assignments, the success rate of the agents rose from 13% to 35% during the same timeframe. These figures suggest that the systems are becoming more reliable at handling sustained, multi-step workflows.

While the overall trend is positive, the results were not uniform across every type of work. The data indicates that the success of these autonomous agents depends heavily on the specific nature of the assignment. Some tasks are currently much easier for the AI to handle than others, creating a varied landscape of reliability for different types of research.

Task-specific performance

The data reveals that agents perform differently depending on whether the work is communicative, analytical, or technical. From August to October 2026, the success rates for communication assignments reached 69%, while analytical tasks saw a 58% success rate. Building research infrastructure was slightly lower, coming in at 43%.

Software development remains a more difficult hurdle for these autonomous systems. The success rate for autonomously completing software for inference and serving systems reached only 25%. OpenAI noted that communication and analytical work generally produced better results than the more complex requirements of software development.

The company also highlighted specific successes in monitoring environments. For example, agents achieved a 79% success rate when tasked with monitoring training runs. However, the results were considerably lower when the agents were asked to actually launch those same training runs, illustrating the gap between observation and execution.

Rising costs of coding

As the capabilities of these agents grow, so does the financial investment required to run them. OpenAI estimated that the median daily spending on coding agents reached $2,400 in October 2026, which represents a 62% increase over the $1,500 spent in September. This trend shows a rapid scaling of how much capital is being funneled into automated coding tasks.

The spending gap between average users and high-volume users is also widening significantly. At the 90th percentile, daily spending on coding agents reached $23,000 in October. This figure is more than double the $11,000 spent at that same percentile during the previous month. These costs reflect the increasing intensity of the work being performed by the agents.

Despite the high costs, OpenAI notes that these coding agents are already providing tangible speed benefits. The systems have reduced certain workflows that previously took several days down to a single day. This acceleration allows researchers to test experimental tools much faster than they could with manual processes alone.

The limits of automation

Even with these improvements, OpenAI emphasizes that autonomous agents are not yet ready to replace human researchers entirely. The systems still face significant hurdles regarding research judgment, software quality, and experiment planning. They often struggle to produce reliable results without constant human intervention to ensure accuracy.

A major concern for the company is the presence of mistakes in assignments that appear to be completed. Because these errors can be subtle, the agents often require close supervision to ensure the output is actually usable. Human researchers continue to provide the essential direction and judgment required to navigate the complexities of the development process.

The company also noted that while measurable research activity is increasing, these figures do not independently prove that scientific progress is moving faster. To track model improvements, OpenAI uses a capabilities index that combines results from multiple AI benchmarks. This index helps them estimate general performance gains even when the raw volume of work doesn't tell the whole story.

Safety and alignment focus

Alongside these performance gains, OpenAI has seen a massive shift in its internal research priorities toward safety. Documented contributions to Safety and Alignment increased by 77% between January and September 2026. This represents the largest absolute increase among all of the company's research categories.

The company attributed this surge in safety work to two main factors: new recruitment and a shift in focus from existing staff. Employees are increasingly moving their attention toward safety-related assignments as the models become more autonomous. This shift suggests that as agents do more work independently, the need for robust safety frameworks becomes more urgent.

The growth in safety research occurs in tandem with the expansion of agentic capabilities. As agents take on more research infrastructure and analytical tasks, the company is prioritizing the alignment of these systems. This dual focus on capability and safety defines the current trajectory of OpenAI's internal development and research goals.

What to watch

  • Whether the success rate for software launching catches up to the success rate for monitoring.
  • How the $23,000 daily spending at the 90th percentile impacts the long-term sustainability of agentic workflows.
  • The specific types of 'mistakes' that currently require the most substantial human correction.

Try it

Look for 'coding agents' in your workflow to see if they can reduce a multi-day task into a single day of work.

Sources and links

Read the original at Mezha.net

First reported by Mezha.net. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news