Agents
OpenAI estimates its agents finish 46% of delegated two-to-four-hour research tasks unaided, up from 23% in March
OpenAI’s November 18 estimate covers delegated tasks and rests on its own methodology; humans remain central, it says.
OpenAI said in a November 18 post titled “Research acceleration: The jagged frontier” that its coding agents completed an estimated 46% of delegated research tasks rated at two to four hours of human work without intervention in October, up from 23% in March, per AI Weekly. The figure is an OpenAI estimate for delegated tasks, not a measure of all researcher work.
The essentials
- Headline figure: 46% for two-to-four-hour tasks in October against 23% in March, per TokenPost.
- Longer tasks: 35% for four-to-eight-hour tasks against 13% over the same period, per TokenPost.
- An uneven picture: between August 1 and October 18, agents succeeded on 79% of tasks monitoring training runs and 35% of those launching them, per AI Weekly.
- Spending: median daily agent use per researcher was valued at $2,400 at API list prices in October, up 62% from September, per TokenPost.
How it was measured
The analysis excludes known automation patterns and gives each researcher-month equal weight; for tasks with unclear outcomes and no recorded intervention it assumes half succeeded, per Kingy AI. Digg notes that changing delegation choices may also affect the results.
Humans stay in the loop
OpenAI emphasises that humans remain central to the research loop, per Phemex. Researchers also reported bad code and faulty analysis that required close oversight, per Digg.
What is not established
- How the 46% would change under a different method: not stated in the sources.
- Whether other labs report comparable figures: not stated in the sources.
- The share of all researcher work that agents now do: not stated in the sources.
Summary table
| Point | What is stated as of November 19, 2026 | Source |
|---|---|---|
| 2–4 hour tasks, March | 23% | TokenPost |
| 2–4 hour tasks, October | 46% | TokenPost |
| 4–8 hour tasks, March to October | 13% to 35% | TokenPost |
| Median daily agent use | $2,400 per researcher | TokenPost |
FAQ
Do agents now do 46% of OpenAI researchers’ work?
No. It is an estimated success rate on delegated tasks rated at two to four hours, per Digg.
Is the figure independently verified?
It comes from OpenAI’s own report; no outside audit is cited.
Method and limits
Checked on November 19, 2026 against AI Weekly, TokenPost, Kingy AI, Digg and Phemex.
Sources
- AI Weekly: OpenAI Says Coding Agents Finish 46% of 2-4 Hour Research Tasks (verified 2026-11-19)
- TokenPost: OpenAI Coding Agents Post 46% Success Rate on Two-to-Four-Hour Tasks (verified 2026-11-19)
- Kingy AI: How OpenAI uses AI to build AI, and where its agents struggle (verified 2026-11-19)
- Digg: OpenAI says its coding agents finish longer research tasks—and still write ‘AI slop’ (verified 2026-11-19)
- Phemex: OpenAI Agents Double Research Task Success Rate to 46% in October (verified 2026-11-19)
Spotted an error in this article? Report it · Our corrections policy