Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days

OpenAI's Own Agents Now Finish 46% of Two-to-Four-Hour Research Tasks Unaided, Up From 23% in March

OpenAI's Own Agents Now Finish 46% of Two-to-Four-Hour Research Tasks Unaided, Up From 23% in March

OpenAI's second research acceleration report says coding agents completed 46% of two-to-four-hour research tasks without human help in October, up from 23% in March, and 35% of four-to-eight-hour tasks, up from 13%. Median agent use reached about 2,400 dollars a day at API list prices, up 62% from September. Documented safety and alignment workweeks rose 77% from January to September. The figures are OpenAI's own internal measurements.

OpenAI published a second report on how its own researchers use coding agents, and the headline number is a doubling: for tasks that would take a human two to four hours, agents now finish 46% without any human intervention, up from 23% in March. For tasks estimated at four to eight hours, the autonomous success rate rose from 13% to 35%. The report, titled "Research acceleration: The jagged frontier" and released on November 18, extends a September baseline with new measures of research inputs, process and output.

The aggregate numbers hide a wide spread. Between August 1 and October 18, agents succeeded unaided on 79% of tasks that involve monitoring a training run for problems, and on 72% of technical questions. They managed only 35% of training launches, 29% of work building evaluation, reinforcement learning and agent environments, and 25% of writing code for inference and serving stacks. Grouped under Epoch AI's O*NET-derived R&D taxonomy, agents scored 69% on communication work, 58% on analysis and 43% on building.

Usage is climbing faster than capability. At API list prices, the median researcher's daily agent consumption reached about 2,400 dollars in October, up 62% from 1,500 dollars in September, while the 90th percentile hit 23,000 dollars, up 105% from 11,000 dollars. OpenAI also counts "agent workdays," where eight hours of agent work equals one day. That measure rose from 0.5 per researcher per day to 5.7, an elevenfold jump, with pretraining and posttraining teams averaging roughly eight agent workdays per researcher per calendar day and scaling and infrastructure teams about four. By mid-October, more than two-thirds of researchers in the analysis averaged a daily peak of at least four concurrent agents, and more than one in ten averaged at least sixteen.

The spending is large enough to change how the lab accounts for research inputs. OpenAI says API-equivalent agentic spending inside the research organisation was about twice non-equity compensation as of September 2026, even though equity still makes up 70% to 80% of target researcher pay and internal deployment remains roughly 2% of total research compute.

Where humans went also shifted. Documented Safety and Alignment workweeks rose 77% from January to September, the largest absolute increase of any research area, driven by both hiring and existing staff moving toward safety work. OpenAI says the growth accelerated after its response to the July 2026 Hugging Face incident and subsequent work on safety cases for frontier training.

A survey of 101 research staff, run from late August to early September, asked what they would trade to keep frontier coding agents if access disappeared for three months. The median answers were 30% of the team's compute, or in a separate question 40% of its headcount. Respondents also said their teams would need a median of three months to reproduce a month of work using only Copilot and ChatGPT. In a later survey, 27 of 28 respondents chose or leaned toward the latest internal model over a mid-career human engineer equipped with a 2024 model.

On output, OpenAI says lines of code changed per active contributor kept rising sharply in 2026, and logged experiments per active experimenter averaged 1.68 times the 2025 baseline in the four weeks ending October 18. The company also introduced an internal index, OECI, modelled on Epoch's Capabilities Index; OpenAI models gained 16.4 points per year on it since August 2025, against 14 points per year on the public ECI since late 2024. OpenAI notes its models score relatively higher on its own mixture than on Epoch's, which it attributes to differences in the benchmark set.

Two caveats run through the document. The first is that rising success rates do not measure the models alone, because researchers may be getting better at choosing which jobs to hand off. The second is that the uplift is jagged rather than uniform, and that agents still rate poorly on research taste, code quality, data quality and writing in the ten interviews OpenAI conducted. OpenAI says it is publishing the numbers because labs should be required to track and report indicators of recursive self-improvement, while stating that the report should not be read as an intention to pursue full RSI. All figures are the company's own internal measurements and have not been independently audited.

For teams outside the lab, the practical reading is narrower than the headline. Delegation works today for supervision and question-answering workloads, and still fails often enough on infrastructure code that the last hour of a task remains human work. OpenAI's own framing is that the weak categories, not the strong ones, will set the ceiling on how much of a research pipeline can be automated.

Comments (0)

Log in to join the discussion

Log In

No comments yet