AI · Nov 9, 2026
OpenAI’s Control Room puts a fail-closed kill switch inside the inference engineOpenAI’s second self-acceleration report: agents now do half the work on mid-length research tasks
· Updated · Signal Desk newsroom
Analysis of reporting by OpenAI. How we work

Agent success on two-to-four-hour research tasks doubled to 46% between March and October, and the median researcher burns $2,400 a day in agent compute. OpenAI frames the uneven gains as reassurance, but the trend lines point the other way.
Key takeaways
- OpenAI estimates agent success on two-to-four-hour research tasks rose from 23% in March to 46% in October, and on four-to-eight-hour tasks from 13% to 35%.
- Success is uneven: 79% for monitoring training runs, 35% for launching them, 25% for writing inference and serving code.
- Median researcher agent use hit $2,400 a day at API list prices in October, with the 90th percentile at $23,000; agent workdays per researcher rose elevenfold to 5.7.
- Agent use is about twice research non-equity compensation in API-equivalent value, but only 2% of total research compute.
- Safety and Alignment documented workweeks rose 77% from January to September, accelerating after the July Hugging Face incident.
The numbers OpenAI chose to publish
OpenAI’s second report on research acceleration is a rare thing: a lab publishing internal metrics on how much its own AI speeds up building the next AI. Its first report, which we covered as the view inside OpenAI, claimed an “automated research intern.” This one adds harder numbers. Estimated agent success on research tasks that take a human two to four hours doubled from 23% in March to 46% in October. For four-to-eight-hour tasks, it rose from 13% to 35%.
Usage is exploding alongside. At API list prices, the median researcher used about $2,400 a day of agent tools in October, up 62% from September; the 90th percentile used $23,000, up 105%. Measured as “agent workdays”, eight hours of agent runtime, researchers went from 0.5 to 5.7 per day, an elevenfold jump, with pretraining and post-training teams near eight. More than two-thirds of included researchers averaged a daily peak of at least four concurrent agents. Asked what they would trade to keep frontier coding agents for three months, the median answers were 30% of team compute or 40% of team headcount.
Jagged is the headline, steep is the story
OpenAI leans on the word “jagged.” Agents succeed 79% of the time at monitoring training runs, 35% at launching them and 25% at writing inference and serving code, so humans remain central. That’s true today. But the report’s own trend lines tell a sharper story. Mid-length task success doubled in seven months. Spending on agents is already about twice the research organization’s non-equity compensation in API-equivalent terms, though equity, 70 to 80% of researcher pay, is excluded, and internal use is only 2% of research compute. And OpenAI says its internal capability index for its models has risen 16.4 points a year since August 2025, faster than the 14 points a year Epoch observed across the field.
If those curves hold, “jagged” describes a snapshot, not a ceiling. The survey answer is telling: researchers would give up 40% of their colleagues before giving up the agents. That’s a strong signal about where labor in frontier labs is heading, even if OpenAI doesn’t say so directly.
Safety effort is growing, but not as fast
The good news is that documented Safety and Alignment effort grew more than any other area in absolute terms, up 77% in workweeks from January to September, from new hires and researchers shifting focus. OpenAI says growth picked up after its response to the Hugging Face incident in July 2026. Still, 77% over nine months is modest next to an elevenfold rise in agent workdays. The report lists differential acceleration of capabilities versus safety as an area it hopes to report on. That’s the number that matters most, and it isn’t here yet.
We credit OpenAI for publishing this at all, and for arguing, as in its frontier standards proposal, that labs should be required to report indicators of recursive self-improvement. It also cites similar reports from Anthropic, Google DeepMind and METR. But self-reported metrics with internal definitions are hard to audit. A footnote notes an August harness change that depressed measured token spend until it was reverted, a reminder of how fragile these measures are. OpenAI says the report shouldn’t be read as an intention to pursue full recursive self-improvement. The data shows it is moving along that path anyway.
Who should pay attention
- AI researchers
- Agent-heavy workflows are now standard at OpenAI; skills shift toward directing and checking many concurrent agents.
- Policymakers
- A lab is publishing recursive self-improvement indicators voluntarily; the case for mandatory, audited reporting gets stronger.
- Safety community
- Alignment effort is rising, but the capabilities-versus-safety comparison is still missing.
- Rival labs
- Pressure to publish comparable metrics, and evidence that agent spend is now a core research input.
What to watch
- The promised report on differential acceleration of capabilities versus safety research.
- Whether task success on four-to-eight-hour work crosses 50% in the next update.
- Any independent audit or external verification of these internal metrics.
- Changes in OpenAI research hiring as agent spend overtakes compensation.
- Similar disclosures from Anthropic, Google DeepMind and xAI.
FAQ
How much do agents do at OpenAI?
OpenAI estimates agents complete about 46% of two-to-four-hour research tasks on their own as of October, with success varying widely by task type.
How much do researchers spend on agents?
The median researcher used about $2,400 a day at API list prices in October, and the 90th percentile about $23,000.
Is OpenAI pursuing recursive self-improvement?
OpenAI says its reporting shouldn’t be read as an intention to pursue full recursive self-improvement, but it argues labs should report indicators of it.
Did safety work grow?
Yes. Documented Safety and Alignment workweeks rose 77% from January to September, according to the report.
Companies: OpenAI
More on OpenAI
AI · Nov 11, 2026
OpenAI’s $2 million for children’s AI safety research matters for one promise: publishing bad newsAI · Nov 17, 2026
OpenAI says its models now shave area off production chips, including AMD’s next flagshipsAI · Nov 3, 2026
OpenAI’s GPT-5.6 Cyber reaches AWS GovCloud through Daybreak Red, with two approval gatesAI · Oct 29, 2026
OpenAI’s new Model Spec swaps rules for five values, trading auditability for judgment