OpenAI Agents Hit 46% Success on Research Tasks

OpenAI reported coding agents completed two-to-four-hour research tasks autonomously 46% of the time in October, up from 23% in March.

Updated on Nov. 19, 2026 in Artificial Intelligence

OpenAI Agents Hit 46% Success on Research Tasks

Live Poll

Do you plan to rely more on AI tools for your own work this year?

In October, OpenAI coding agents autonomously completed 46% of research tasks estimated at two to four hours, compared with 23% in March. Success on four-to-eight-hour tasks reached 35% in October, up from 13% in March.

Why it matters

Rising autonomous success on longer tasks signals that AI agents are taking on more substantive research work, which OpenAI said prompted a call for labs to track and publicly disclose recursive self-improvement measures.

Autonomous success on two-to-four-hour research tasks rose to 46% in October, while four-to-eight-hour tasks reached 35%, leaving an 11-percentage-point gap between the bands. The extraction does not describe the task set or independent verification of these figures.

The players

OpenAI

OpenAI is an artificial intelligence research company that develops the GPT family of large language models and the Codex coding agents discussed in this report.

The details

Median daily agent use among researchers reached $2,400 in October at API list prices, up 62% from September, while 90th-percentile daily use hit $23,000, up 105%. Recorded workweeks in safety and alignment rose 77% from January through September.

Timeline

  1. March: baseline coding agent success rates recorded for both task bands.

  2. January through September: safety and alignment workweeks increased 77%.

  3. September: baseline month for the 62% rise in median daily agent use.

  4. October: coding agent success rates and agent use valuations recorded.

The Tech Race

The reported gains on longer research tasks place OpenAI's Codex agents against rival coding-agent products in an industry race to automate multi-hour engineering work. The figures show a measurable step in that competition rather than a settled lead.

Researchers and developers who rely on coding agents may see these tools handle longer, more complex tasks with less supervision. The extraction gives no pricing or access changes for end users.

The takeaway

Track how often AI agents complete multi-hour work without supervision, since that measure is becoming a central test of their practical value. Watch for whether labs publish the recursive self-improvement metrics OpenAI has requested.

Further reading

Explore more reporting in our Artificial Intelligence section.

Live Poll

Do you plan to rely more on AI tools for your own work this year?