OpenAI maps research acceleration’s “jagged frontier”: longer-task success rises, with humans still central
U.S. company OpenAI has updated its internal research-acceleration report. Researchers estimate that, from March to October, success without human intervention rose from 23% to 46% for tasks estimated to take humans 2–4 hours, and from 13% to 35% for 4–8-hour tasks. These are not validated, standardized benchmark results, and changing task delegation also affects performance. The report describes growing usage and human–AI collaboration while separating its policy case for tracking recursive self-improvement from measured historical trends.
U.S. company OpenAI released Research acceleration: The jagged frontier on November 18, 2026, updating its research-acceleration report from September this year. Its official account posted the news at 23:07:23 UTC on November 18 (7:07:23 a.m. on November 19, Beijing time). It describes the material as a “broader but still incomplete” internal picture: AI is taking on more research work, but gains are uneven, and humans remain central to research.
The most striking change concerns longer tasks. The report says that, from March to October, estimated success without human intervention rose from 23% to 46% for tasks estimated to require 2–4 hours of human work, and from 13% to 35% for 4–8-hour tasks. Both human task duration and success rates are researcher estimates, not a validated, standardized benchmark. The tasks delegated to agents are also changing, so the gains do not depend solely on any single model becoming stronger.
Differences between tasks matter just as much. The report lists success rates of 79% for monitoring training, 35% for launching training, and 25% for writing code for the inference and serving stacks. Here, the serving stack is the system that organizes the resources needed to run models. These differences form a “jagged” pattern rather than a smooth rise across every part of research.
Usage intensity is increasing too. Valued at API list prices—the prices on the price list, rather than actual bills paid—the reported median daily usage value rose from $1,500 in September to $2,400 in October, an increase the report puts at 62%. The displayed figures are rounded. The increase combines more activity, more expensive models, and changes in model behavior and tool configurations. A temporary efficiency change on August 21 coincided with a decline in input tokens; beginning September 10, that change was gradually rolled back for reasons unrelated to spending, after which recorded spending rose. Assigning the increase to a single cause would leave out this context.
Another measure is “agent workdays,” with every 8 hours of use counting as 1 workday. The report gives a per-person increase from 0.5 to 5.7, roughly elevenfold. Pre-training and post-training teams recorded about 8 per calendar day, and scaling infrastructure teams about 4. The measure primarily captures interactive use and may undercount automation. It measures usage, not an equivalent quantity of human research output.
The origin of ideas has more specific boundaries too. In a small inquiry among key contributors to GPT-6 Astra and 6.2 Astra, both the post-training and safety groups attributed the named ideas to humans, with models helping implement them. Pre-training responses likewise credited humans for 6 Astra; for 6.2 Astra, some contributors developed ideas in conversations with Codex. This limited account does not cast models as the source of all core ideas.
The report also includes a questionnaire of 101 people. Under the assumption that no coding agent stronger than GPT-5.2 would be available over the next 3 months, respondents’ stated resource trade-offs to retain agents involved 40% of team headcount or 30% of team compute. These were hypothetical responses, not layoff plans or actual transactions. They express the value respondents perceived in the tools, rather than resource substitution already realized.
The policy argument is a separate layer. OpenAI argues that AI laboratories should track and publish indicators of RSI—recursive self-improvement, meaning AI accelerating the research and development of future AI systems—to inform governance discussions about more powerful AI. This is the company’s position, not a rule already in force. The report also says it should not be read as an intention to pursue full RSI. Historical trends involving substantial human–AI collaboration may not predict future AI-driven progress. It is monitoring whether deep learning might experience a sudden breakthrough similar to recent developments in mathematics, rather than reporting that such a deep-learning breakthrough has already been found.
For ordinary readers, the material offers a window into changing research work: AI can take on longer tasks while still performing very differently across activities. Greater usage and researchers’ increased willingness to rely on tools are separate questions from whether research output grows by the same amount.
iHolyshit review
I find the combination of rising long-task success and persistent gaps between tasks most meaningful. It makes AI’s changing role in research concrete while reminding me that humans framing problems, choosing what to delegate, and driving implementation remain important parts of the picture. Keeping usage growth, perceived value, and research progress separate is more useful than compressing every change into a single acceleration multiplier.
Share this article
What do you think of this story?
Anonymous feedback, up to 5 reactions per article.
Reader comments
Comment without signing in, or sign in to display your member name.
Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.
No comments yet. Be the first to share.