Skip to content
The Dragon's Web

News Portal

The Dragon's Web

News Portal

  • Home
  • Local Weather
  • Home
  • Local Weather
Close

Search

OpenAI claims AI agents now complete nearly half of two-to-four-hour research tasks on their own – but says ‘humans are still central to AI development’
ProTech News

OpenAI claims AI agents now complete nearly half of two-to-four-hour research tasks on their own – but says ‘humans are still central to AI development’

By TechRadar
November 22, 2026 3 Min Read
Comments Off on OpenAI claims AI agents now complete nearly half of two-to-four-hour research tasks on their own – but says ‘humans are still central to AI development’
  • AI agents now complete 46% of two- to four-hour research assignments
  • Success rates for longer research assignments have more than doubled since March 2026
  • Agents perform best in communication tasks and struggle with technical research responsibilities

OpenAI says its AI agents are becoming increasingly capable of handling research assignments, including work previously requiring several hours of human effort.

The company’s latest internal assessment indicates that autonomous task completion is improving, although performance varies considerably depending on the nature of the assignment.

Despite these advances, OpenAI maintains that human researchers continue to provide essential direction, supervision, and judgment throughout its development process.

AI agents are taking on increasingly demanding research assignments

According to OpenAI, the proportion of research tasks completed without human intervention increased substantially between March and October 2026.

Agents independently completed 46% of assignments estimated to require two to four hours of human work, compared with just 23% in March.

For assignments requiring four to eight hours, their estimated success rate increased from 13% to 35% during the same period.

However, these improvements do not translate equally across every research activity, with some technical responsibilities remaining particularly difficult for autonomous systems.

OpenAI found that agents achieved a 79% success rate when monitoring training runs, while launching those runs produced considerably lower results.

Writing software for inference and serving systems proved more difficult, with autonomous completion reaching just 25%.

Performance also varied across research categories, with communication and analytical assignments generally producing better results than software development.

Between August and October 2026, agents achieved estimated success rates of 69% for communication tasks and 58% for analytical assignments.

Building research infrastructure recorded 43%, indicating that substantial differences remain across the technical responsibilities involved in developing advanced AI.

Meanwhile, spending on coding agents has increased significantly among researchers, particularly those making the heaviest use of available systems.

OpenAI estimates that median daily expenditure reached $2,400 in October, representing a 62% increase from September’s $1,500.

At the 90th percentile, daily spending climbed to $23,000, more than doubling September’s $11,000 figure.

Human researchers remain important despite growing AI capabilities

OpenAI’s findings also indicate that researchers are spending more time addressing safety concerns associated with increasingly capable AI tools.

Documented contributions to Safety and Alignment increased 77% between January and September, representing the largest absolute increase among its research categories.

The company attributed this growth to recruitment and to existing employees shifting their attention toward safety-related assignments.

Coding agents have made it much faster for researchers to test experimental tools, reducing work that previously took several days to just one day.

However, AI systems still struggle with research judgment, software quality, experiment planning, and producing reliable results.

These agents frequently require close supervision because apparently completed assignments can contain mistakes requiring substantial human correction.

OpenAI also reported increases in measurable research activity, although these figures do not independently establish faster scientific progress.

The company compared its models’ progress using a capabilities index that combines results from multiple AI benchmarks to estimate improvements in general model performance.

Its internal index, which uses both public and company-developed tests, recorded an annual increase of 16.4 points since August 2025, compared with the 14-point annual increase measured by Epoch AI since late 2024.

However, the two measurements use different combinations of benchmarks, making the figures an indication of progress rather than a direct comparison of identical evaluations.

Google logo on a black background next to text reading 'Click to follow TechRadar'

Go to story

Tags:

111131416223254AagentsAIAI agentsAI toolscodingemployeesHumansinferenceinfrastructuremeasurementsmodelsOpenAIProresearchSafetysoftwarespendingtimeTrainingWork
Author

TechRadar

Follow Me
Other Articles
What “Looney Tunes” Villain Are You?
Previous

What “Looney Tunes” Villain Are You?

A 13-year-old dies after participating in a Benadryl TikTok ‘challenge’
Next

A 13-year-old dies after participating in a Benadryl TikTok ‘challenge’

About This Site

This is my personal site I use to test code, practice techniques, and show off my skills.

Recent Posts

  • Two stars in the Andromeda galaxy merged before astronomers’ eyes
  • Oakdale Police Beat 11-18-25
  • Knights Ferry welcomes Salmon Festival visitors
  • An El NiƱo amped by climate change has California bracing for the worst
  • Earth just saw the most anomalously hot day ever on record

Archives

  • November 2026 (766)

Find Us

Address
1460 Grand Oak Way
Oakdale, CA 95361

Copyright 2026 — The Dragon's Web. All rights reserved. Blogsy WordPress Theme

Powered by CookieAdmin
Necessary cookies enable essential site features like secure log-ins and consent preference adjustments. They do not store personal data.
None
Functional cookies support features like content sharing on social media, collecting feedback, and enabling third-party tools.
None
Analytical cookies track visitor interactions, providing insights on metrics like visitor count, bounce rate, and traffic sources.
None
Advertisement cookies deliver personalized ads based on your previous visits and analyze the effectiveness of ad campaigns.
None
Unclassified cookies are cookies that we are in the process of classifying, together with the providers of individual cookies.
None
Powered by CookieAdmin