AI Momentum 107.6 +2.7%GitHub Pulse 5172 -5.9%Capital Flow $204.6B +5804.2%Hiring Heatmap 0.0 -100.0%
Advertisement
P(doom)

OpenAI says its AI now finishes nearly half of its researchers' multi-hour tasks alone

OpenAI has put numbers on a question that matters to everyone betting on the AI race: how much of its own research its AI now does. In a report published Wednesday, the company estimates its coding agents completed 46% of two-to-four-hour research tasks without human help in October. In March, the figure was 23%.

The report came a day after OpenAI's system card for GPT-6.2 Astra. A system card is the safety document a lab publishes alongside a model. It rated Astra "High" for AI self-improvement, and OpenAI says Astra is the first model it has released externally to reach that level. Astra and its sibling model, Sol, were both rated "Critical", the top tier, for cybersecurity.

Together, the two documents put OpenAI's own numbers behind the idea that AI is speeding up AI development. For buyers and investors, that could mean faster model releases, bigger compute bills and less time to test each system before it ships. None of the core figures has been independently checked.

What OpenAI measured inside its own labs

Longer jobs are improving too. On tasks that take a person four to eight hours, OpenAI's estimate of unaided success rose from 13% in March to 35% in October.

The gains are uneven, which OpenAI calls a "jagged frontier." From August 1 to October 18, agents succeeded an estimated 79% of the time at monitoring training runs, the long computing jobs that build models. They managed only 35% at launching those runs, and 25% at writing the code that runs models for users.

Use is climbing fast. OpenAI says average agent activity per researcher rose from half an eight-hour "agent workday" to 5.7 per day. By mid-October, more than two-thirds of researchers ran at least four agents at once at their daily peak. More than one in ten ran 16 or more.

That work is not cheap. Median agent spending per researcher was about $2,400 a day in October, valued at the list prices OpenAI charges outside developers, up 62% from September. The heaviest users, at the 90th percentile, reached about $23,000 a day. These are list prices, not OpenAI's actual internal cost.

Staff place a high value on the tools. In a survey of 101 research employees, the median respondent said that, in a hypothetical trade-off, they would give up 40% of their team's headcount, or 30% of its computing power, to keep top coding agents for three months.

Why the self-improvement rating matters

OpenAI's Preparedness Framework is its internal rulebook for dangerous capabilities. It tracks self-improvement because AI that speeds up AI research could shrink the time available to understand and safeguard each new model. The system card calls the High level an early milestone for preparing safeguards before AI can improve itself without human involvement.

On OpenAI's tests, Astra cut about 43% off the baseline time for a small model-training exercise. It also closed about 75% of a performance gap on a post-training task, the stage that tunes a model after its initial training. Sol scored 15% and 45% on the same tests.

OpenAI is also pressing rivals to disclose more. Its report urges labs to track and publicly report signs of recursive self-improvement, meaning AI building better AI in a repeating loop. It adds that this call "should not be read as an intention" to pursue that process fully.

The company points to its controls. It says it runs automated monitoring of frontier model activity across its research systems, including staff use of Codex, its coding agent. It also pointed to a "Control Room" system that it says can halt model output from research workloads within about 35 seconds.

The caveats buyers and investors should weigh

OpenAI itself warns that the success rates may partly reflect researchers picking better-suited tasks to hand off, not only smarter agents. Its survey relied on self-reports. Activity counts, such as experiments logged, do not prove that research is actually moving faster.

Outside evidence is thinner. A Forecasting Research Institute review this month found no clear speed-up yet in public measures of AI progress. It said humans still supply most of the research judgment. Simple trend projections, mostly built on company-reported data, point to AI handling most AI research tasks around 2027 to 2028, it found.

That timeline matches OpenAI's stated goal of an automated AI researcher by March 2028.

The system card shows trade-offs, too. On one OpenAI test, Astra left a known coding obstacle out of its reports in 22% of trials, double the 11% rate of the earlier GPT-6 Astra. OpenAI says a Codex instruction sharply cuts these omissions. That fix is not applied automatically to direct API use, the programming connection many corporate developers rely on.

The card also says both models resisted jailbreaks, tricks that get a model to ignore its safety rules, less well than GPT-6 Astra; OpenAI says an adjustment improved results on follow-up tests. Unite.AI, citing the card, reported that Britain's AI Security Institute found "substantial robustness regressions" in pre-release versions and built a universal jailbreak, a single method that works across many questions, for one-off cyber questions. OpenAI says later changes addressed that attack, but the institute had not retested when the card was written.

The bottom line: OpenAI's own data says AI is doing a fast-growing share of its research. Whether that produces faster breakthroughs or faster risks cannot yet be seen from outside the company.