report · 2026-09-06

Research acceleration: The view inside OpenAI

OpenAI

Why it matters here

OpenAI’s internal charts show improving success on researchers’ tasks alongside a continuing need for human intervention. Together, they help distinguish useful research assistance from research that runs without human steering.

Observed R&D automationAI-R&D capability

What the charts show

More research tasks completed without intervention

OpenAI reports that success without human intervention generally increased across several task-duration groups from January to July 2026. This shows growing ability to carry out parts of researchers’ work.

OpenAI Research coding-agent sessions, January–July 2026. Time buckets estimate how long a human would need, not how long the agent ran. Results give each user equal weight and exclude uncertain outcomes.

Monthly success curves with 95% bootstrap confidence intervals, grouped by estimated human task duration. Later months generally show higher success; some long-duration cells are absent. — OpenAI source chart rendered as an image, with its original axes, plotted marks and legend. No values inferred or redrawn. Enlarge chart ↗

View original chart and methods ↗
Research acceleration: The view inside OpenAI · Section 3, “Agents are increasingly solving more complex tasks for researchers”; View methods

How to interpret this chart

The result is successful tasks with zero interventions among eligible tasks, not success among only the tasks that received no intervention.
Outcomes and human-effort estimates were classified by GPT-5.6 Sol; this identifies the evaluator, not the models being evaluated. Human agreement was checked on 25 tasks, without a published accuracy rate.
Task mix, users, models and software changed over time. Duration grouping only partly controls these differences; this is not a fixed benchmark or a causal measure of research productivity.
The caption excludes cells with fewer than 50 sessions or 50 users; the accessible description calls these tasks. Each sampled session is assigned one primary task. Late-July sessions lack a full seven-day follow-up.

Published values from the chart

All 57 displayed point labels, at the source’s published precision. Half-widths are copied from accessible labels; missing or suppressed cells are not zero. No unrounded estimates or inferred interval endpoints are supplied.

More research tasks completed without intervention
Month	Estimated human time	Success without intervention	95% interval half-width
2026-01	<15m	63%	7.6%
2026-01	15-30m	62%	10.1%
2026-01	30m-1h	56%	7.1%
2026-01	1-2h	51%	7.3%
2026-01	2-4h	28%	6.4%
2026-01	4-8h	18%	6.7%
2026-01	8-16h	10%	8.0%
2026-02	<15m	74%	4.9%
2026-02	15-30m	76%	6.5%
2026-02	30m-1h	62%	5.3%
2026-02	1-2h	52%	5.0%
2026-02	2-4h	43%	5.3%
2026-02	4-8h	35%	6.3%
2026-02	8-16h	24%	8.3%
2026-03	<15m	80%	3.6%
2026-03	15-30m	77%	5.1%
2026-03	30m-1h	68%	4.3%
2026-03	1-2h	59%	3.8%
2026-03	2-4h	58%	3.9%
2026-03	4-8h	48%	5.1%
2026-03	8-16h	23%	5.9%
2026-03	16-32h	20%	8.5%
2026-04	<15m	91%	1.8%
2026-04	15-30m	81%	4.3%
2026-04	30m-1h	64%	3.9%
2026-04	1-2h	64%	4.0%
2026-04	2-4h	58%	3.9%
2026-04	4-8h	39%	4.3%
2026-04	8-16h	36%	5.6%
2026-04	16-32h	22%	7.5%
2026-04	32-64h	18%	8.3%
2026-05	<15m	90%	1.9%
2026-05	15-30m	77%	4.6%
2026-05	30m-1h	72%	3.7%
2026-05	1-2h	62%	3.9%
2026-05	2-4h	62%	3.7%
2026-05	4-8h	43%	4.6%
2026-05	8-16h	43%	5.2%
2026-05	16-32h	24%	8.8%
2026-05	32-64h	8%	6.7%
2026-06	<15m	81%	3.2%
2026-06	15-30m	77%	4.7%
2026-06	30m-1h	75%	3.6%
2026-06	1-2h	68%	4.0%
2026-06	2-4h	58%	4.1%
2026-06	4-8h	54%	4.9%
2026-06	8-16h	58%	4.8%
2026-06	16-32h	30%	10.4%
2026-07	<15m	87%	2.1%
2026-07	15-30m	76%	4.4%
2026-07	30m-1h	75%	3.4%
2026-07	1-2h	64%	4.0%
2026-07	2-4h	57%	4.1%
2026-07	4-8h	53%	4.7%
2026-07	8-16h	35%	5.8%
2026-07	16-32h	35%	8.6%
2026-07	32-64h	17%	7.7%

Longer tasks still need human steering

OpenAI reports that more than half of successful tasks in its four-to-eight-hour category needed human intervention. Better task performance therefore still comes with a substantial role for people.

Stacked outcome bars by estimated human task duration; the share of successful tasks needing intervention is larger in longer-duration groups. — OpenAI source chart rendered as an image, with its original axes, plotted marks and legend. No values inferred or redrawn. Enlarge chart ↗

View original chart and methods ↗
Research acceleration: The view inside OpenAI · Section 3, “Longer tasks need more interventions”; View methods

How to interpret this chart

Bars partition non-uncertain tasks into success without intervention, success with intervention, failure, tool errors and no clear goal. They do not show percentages among successful tasks alone.
The chart pools January–July 2026; the accompanying “over half” statement refers to the last six months. Keep that statement separate from the pooled chart.
Interventions mean a person corrected or redid work, including in later sessions. Asking for an expanded or follow-on task does not count.
Internal, classifier-assessed outcomes are not independent replication or a measure of fully autonomous research.

Published values from the chart

All 50 displayed category labels. Shares are averaged across users and rounded by the source, so printed categories may not add to exactly 100%. A printed 0% is a rounded label, not proof that no such cases occurred.

Longer tasks still need human steering
Estimated human time	Outcome	Share of non-uncertain tasks
<15m	Success + 0 interventions	86%
<15m	Success + ≥1 intervention	8%
<15m	Failure	3%
<15m	Tool errors	1%
<15m	No clear goal	2%
15–30m	Success + 0 interventions	76%
15–30m	Success + ≥1 intervention	19%
15–30m	Failure	5%
15–30m	Tool errors	1%
15–30m	No clear goal	0%
30m–1h	Success + 0 interventions	69%
30m–1h	Success + ≥1 intervention	23%
30m–1h	Failure	6%
30m–1h	Tool errors	1%
30m–1h	No clear goal	0%
1–2h	Success + 0 interventions	61%
1–2h	Success + ≥1 intervention	29%
1–2h	Failure	7%
1–2h	Tool errors	2%
1–2h	No clear goal	0%
2–4h	Success + 0 interventions	57%
2–4h	Success + ≥1 intervention	33%
2–4h	Failure	8%
2–4h	Tool errors	2%
2–4h	No clear goal	0%
4–8h	Success + 0 interventions	43%
4–8h	Success + ≥1 intervention	45%
4–8h	Failure	10%
4–8h	Tool errors	2%
4–8h	No clear goal	0%
8–16h	Success + 0 interventions	40%
8–16h	Success + ≥1 intervention	48%
8–16h	Failure	11%
8–16h	Tool errors	1%
8–16h	No clear goal	0%
16–32h	Success + 0 interventions	23%
16–32h	Success + ≥1 intervention	59%
16–32h	Failure	15%
16–32h	Tool errors	3%
16–32h	No clear goal	0%
32–64h	Success + 0 interventions	13%
32–64h	Success + ≥1 intervention	63%
32–64h	Failure	13%
32–64h	Tool errors	10%
32–64h	No clear goal	0%
64–128h	Success + 0 interventions	16%
64–128h	Success + ≥1 intervention	51%
64–128h	Failure	26%
64–128h	Tool errors	7%
64–128h	No clear goal	0%

What to keep in mind

Company-reported operational evidence, not a standardized cross-lab benchmark or an overall RSI score.
A random 2% of conversations was sampled, restricted to interactive Research coding-agent sessions. Automations, memory updates, title generation and ambient suggestions were excluded. The research organization includes infrastructure and project-support roles.