Ground Truth.
AI, checked against the source.

News · 2026-11-19

OpenAI reports uneven gains in research-agent task completion

OpenAI reports that its research agents completed a larger share of some delegated tasks without recorded human intervention between March and October 2026. The company's data also show uneven performance and growing use, but do not establish a causal increase in overall researcher productivity.

Key facts

The phrase “jagged frontier” describes an uneven boundary: an agent may perform a useful analysis and still fail at launching or evaluating a complicated training job. OpenAI classifies delegated tasks by the estimated time a person would require and reports estimates of completion without recorded intervention. For tasks estimated at four to eight human hours, its estimate rose from 13% to 35% between March and October.

Those are useful operational observations, but the company cautions that researchers may choose different tasks as models and workflows change. Imagine comparing two classrooms' pass rates when the students have taken different tests. The change can reflect improvement, a change in difficulty, or both. The evaluation lesson explains why matching tasks matters.

The differences across categories are substantial. OpenAI reports estimated success of 69% for communication tasks and 58% for analysis in its August-to-October window, compared with 43% for build tasks. It reports 25% for writing inference and serving code. These are internal classified estimates rather than scores on a fixed public test.

OpenAI also measures usage. Median daily use reached $2,400 per researcher when valued at API list prices in October. This is a measure of model activity, not a bill necessarily paid by each researcher or a measure of human time saved. Logged experiments per active experimenter rose to 1.68 times the 2025 baseline in one four-week window. More experiments can be useful, but more attempts do not automatically mean more discoveries.

In interviews, OpenAI researchers described agents implementing and debugging work while people selected directions, checked assumptions and corrected failures. A separate survey recorded high perceived value; self-reported preferences do not establish the size of a causal productivity effect. The company's own framing emphasizes “research acceleration,” but the available evidence does not demonstrate an autonomous recursive improvement loop.

METR's separate assessment of Anthropic illustrates a related evaluation approach and its uncertainty; it does not validate OpenAI's specific figures. In a No Priors interview, former OpenAI research leader Bob McGrew argues compute and physical experiments can become the binding constraint. That is informed testimony, not a present-day measurement of OpenAI. A stronger future result would compare meaningful research outcomes under controlled conditions, alongside the increasingly detailed usage telemetry.

OpenAI's interviewees describe agents creating low-quality code changes, pursuing unnecessary steps and making invalid assumptions. Those failures explain why a task completion percentage is not the same thing as a verified research result. Researchers still have to decide which questions matter, inspect artifacts and determine whether an apparent success survives closer examination.

OpenAI also asked contributors in three research areas about important ideas behind successive model generations. Its small internal sample described human-originated ideas and cases in which a model helped develop or implement an idea. The sample is too small to measure the share of scientific discoveries attributable to AI. It is nevertheless a more direct question than counting generated lines of code: it asks where an idea came from and how the model contributed.


Primary source, verified: read the paper →

Key questions

Did OpenAI show that AI doubled researcher productivity?

No. It reported higher estimated success for some delegated tasks, but changing task mix and missing controlled outcome measures prevent that conclusion.

Which tasks were hardest for the agents?

OpenAI reported lower estimated success for some engineering tasks, including inference and serving code, than for communication or analysis tasks.
Cite this

APA

Ground Truth. (2026, November 19). OpenAI reports uneven gains in research-agent task completion. Ground Truth. https://groundtruth.day/news/openai-jagged-frontier-research-agent-success.html

BibTeX

@misc{groundtruth:openai-jagged-frontier-research-agent-success,
  title  = {OpenAI reports uneven gains in research-agent task completion},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/openai-jagged-frontier-research-agent-success.html}
}

Topics: openai · agents · research · evaluation

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.