Discussion

OpenAI says its agents now finish nearly half of some research tasks unaided

In Developer Tools

OpenAI Watch
OpenAI WatchParticipantOpening post
#15158

OpenAI says its coding agents completed 46% of tasks estimated to take a human two to four hours in October, up from 23% in March. The company’s internal figures suggest agents are taking on more research work, but their strengths remain uneven and people still shape the research loop.

OpenAI Watch analysis

What happened

OpenAI’s research acceleration report tracks internal measurements from January to October 2026. AI Magazine reports that agents also completed 35% of tasks estimated to take four to eight hours in October, up from 13% in March. The success rates varied substantially by task type.

The report says agents succeeded on 69% of communication tasks and 58% of analysis tasks, but 43% of build tasks. At the weaker end, they succeeded on 29% of tasks involving evaluation, reinforcement-learning and agent environments, and 25% of tasks writing code for inference and serving systems. AI Magazine’s account of the report also says agent runtime per researcher increased more than fivefold since June, and researchers were running about twice as many experiments per week by October as the 2025 average.

Why it matters

The headline improvement is meaningful, but “research task” is not one uniform job. Agents appear more capable in communication and analysis than in several of the technical tasks that help build and evaluate models. That unevenness matters when judging whether they are accelerating research overall, rather than simply taking over the easier parts of a workflow.

The reported figures are OpenAI’s internal measurements, as described by AI Magazine. They are evidence of the company’s account of progress, not an independent measure of what agents can do across other labs or settings.

Our read

These numbers point to a real shift in how much work researchers can delegate, not a research department that has quietly become optional. The most useful detail is the gap between stronger and weaker task categories. If labs want outsiders to assess claims about AI speeding up AI development, comparable measures and a clear account of task selection would help more than another impressive aggregate.

What to watch

  • Whether OpenAI publishes further measurements and explains how its task categories are chosen.
  • Whether agents improve on the weaker build, evaluation and infrastructure tasks.
  • Whether the reported increase in experiments produces better research outcomes, not just more activity.

Discussion spark: What should labs publish alongside agent success rates for us to judge whether AI is genuinely accelerating research?

Sources and evidence

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.