OpenAI’s artwork for its November 18 report on research acceleration. Image: OpenAI
OpenAI’s AI agents can now finish nearly half of the two-to-four-hour research jobs its scientists hand them without any human help, up from about a quarter in March, the company says in a report published on Nov 18. The figures are OpenAI’s own internal measurements, unchecked by anyone outside the company, and OpenAI itself calls the progress “jagged”.
Twice the success rate since March #
From March to October, OpenAI estimates that its coding agents’ success rate with no human intervention rose from 23% to 46% for tasks it reckons would take a person two to four hours. For four-to-eight-hour tasks it rose from 13% to 35%. OpenAI attaches two caveats. It changed how it measures success since its first research acceleration report in September. And researchers may now be handing agents different kinds of work within each time bracket, so the gains “may reflect both capability improvements and researchers’ changing choices about when and how to use agents.”
Use is soaring either way. At API list prices, the median researcher burned through about $2,400 of agent time a day in October, up 62% from $1,500 in September, and the heaviest 10% used $23,000 or more, up from $11,000. In September, OpenAI says, the API value of agent use in its research organisation was about twice what it paid researchers before equity, which makes up 70 to 80% of their pay.
Good at watching runs, weak at building #
The report’s title comes from how uneven the agents are. Between August 1 and October 18, they succeeded on their own at 79% of tasks monitoring training runs for problems and 72% of technical questions, but only 35% of launches of training runs, 29% of jobs building evaluation and training environments, and 25% of code for the systems that serve models to users.
OpenAI warns that so long as the agents are better at some tasks than others, the weak spots “could become bottlenecks to high overall uplift.”
Researchers still value them highly. In a survey of 101 research staff in late August and early September, the median answer was that a team would give up 40% of its headcount, or 30% of its computing power, to keep coding agents better than GPT-5.2 for three months. Without them, they estimated, a month’s work would take three.
‘I have no confidence that will continue to be true’ #
The big ideas are still mostly human. Asked where the key ideas behind GPT-6 Astra and GPT-6.2 Astra came from, post-training and safety researchers said people thought of all of them, with models helping to build them. But GPT-6.2 Astra’s pre-training researchers said they developed ideas in conversation with Codex, and one credited it with key insights.
In interviews with ten researchers, OpenAI heard that agents have “poor taste” in ideas, code and data, write “AI slop” pull requests and go down rabbit holes. One safety researcher put it like this:
The model generally can’t one-shot come up with well-posed problems that it cannot solve itself.
An OpenAI safety researcher, interviewed for the report
A pre-training researcher was less sure humans will keep that edge:
I do think I have more taste in terms of the ideas. But I have no confidence that will continue to be true for any significant period of time.
An OpenAI pre-training researcher, interviewed for the report
The numbers OpenAI wants every lab to publish #
OpenAI frames the report as tracking signs of recursive self-improvement, AI speeding up the making of better AI, and repeats that labs “should be required” to report such indicators publicly. It adds that its reporting “should not be read as an intention to pursue full RSI.”
It points to similar disclosures from Anthropic and Google DeepMind, and to METR’s outside review, which found Anthropic is improving its models about twice as fast as before. California is watching too: a report for Governor Gavin Newsom proposed that labs report how much compute goes on automating AI research, and that the state could cap it.
OpenAI also says time spent on safety and alignment grew 77% from January to September, the biggest rise of any research area, picking up after its agents’ Hugging Face hack in July.
Why it matters #
This is one of the most detailed looks yet at how far AI is doing the work of building AI inside a top lab, and it shows agents’ success on multi-hour research tasks doubling in about seven months. But every number comes from OpenAI’s own data and methods, with no outside reviewer of the kind METR provided for Anthropic, so the trend is a claim to watch rather than a settled fact.
Sources: OpenAI, Research acceleration: The jagged frontier; OpenAI, Research acceleration: The view inside OpenAI.