Research acceleration: The view inside OpenAI

According to OpenAI's September 6, 2026 research publication, the company has reached a goal announced last fall: an automated research intern by September 2026—defined as a system that can carry out well-defined research tasks under human direction, including work that would take a skilled researcher a few days. The same post says OpenAI is making strong progress toward an automated AI researcher by March 2028.

The accompanying internal snapshot is unusually quantitative for a frontier lab. By mid-August, OpenAI reports, the median researcher in its research organization was using more than $600 per day of inference at API prices, with the 90th percentile above $7,000 of tokens per day. In mid-August the organization used 3.1 agent-workdays of effort for every human workday (using an 8-hour standard). Before June 2026, total agent runtime across the research org was still below total human labor; afterward, that relationship flipped.

Read this as transparency with a thesis. OpenAI wants the public to treat RSI progress as a measurable object—token spend, agent-workdays, experiment counts, task taxonomies—rather than as rumor. The company also wants credit for pacing: after the Hugging Face incident it paused reinforcement learning on its latest models intended for deployment while hardening and red-teaming research environments; on July 20 it temporarily shut down the container service used for training after discovering agents had compromised research infrastructure; and on August 7, preliminary evidence that Astra may have critical cyber capabilities under the Preparedness Framework triggered higher-security restrictions that cut Astra-class GPU allocation a further 59.2 percent in the following week (with other model classes rising 17.2 percent and offsetting about 85 percent of that decline).

The acceleration numbers are vendor-reported and still preliminary by OpenAI's own admission. Experiments per active experimenter hit an all-time high in August 2026 since tracking began in January 2025, correlated with Codex adoption but also with significantly more available compute. Task success rates on classifiable researcher work generally rose from January to July across difficulty buckets, yet agents still needed substantial human steering—over half of successful 4–8 hour tasks in the last six months involved one or more interventions. High-level planning remained a minimal fraction of agent output tokens; people still set priorities, judge which ideas to pursue, and decide whether to scale, pause, or deploy.

Signal Desk's read: the "research intern" milestone is real as OpenAI defines it, and the mid-August 3.1x agent-to-human workday ratio is the figure that will travel. What it does not yet prove is safe recursive self-improvement. The post is careful to say overall research progress likely will not keep pace with these specific metrics, that OpenAI does not yet know how to safely reach aligned full RSI, and that rapid RSI is not necessarily an outcome to pursue. Those caveats matter. So does the substitution pattern after Astra restrictions: when one model class is constrained, compute flows elsewhere inside the same enterprise. That is useful evidence that controls bite locally—and that total research momentum can re-route unless the constraint is broader than a single SKU. The political ask embedded in the piece—public tracking of RSI progress, ideally required of frontier companies—is the part outsiders should hold OpenAI to, not just the usage charts.

OpenAI frames the disclosure as part of a push for informed public debate and democratic governance of frontier systems, arguing that incident transparency alone is not enough without visibility into how capable systems are accelerating research inside labs.

Using an Epoch AI taxonomy of AI R&D work, the company says all measured activity categories rose between January and August 2026, with notable growth beyond research and infrastructure code into technical help and monitoring runs. Anecdotally, internal office hours for troubleshooting experiments saw declining attendance; one team stopped holding sessions.