OpenAI published a detailed self-assessment of how much its own researchers now rely on AI agents, framed explicitly around progress toward recursive self-improvement (RSI) — a system that improves the intelligence used to build its own successor. The headline claim: OpenAI says it has reached a goal it announced last fall, having "an automated research intern" — a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days — in place by September 2026, on schedule. The next target, more ambitious and further out: "an automated AI researcher by March of 2028."
The usage numbers — with one figure that needs a grain of salt
The concrete data is genuinely striking. At the start of 2026, the median researcher at OpenAI used coding agents "only in modest amounts." By mid-August, the median researcher was spending more than $600 a day on agent inference at API prices, and the 90th-percentile researcher was spending more than $7,000 a day. Sometime before June 2026, total agent runtime across the research organization overtook total human labor; by mid-August, OpenAI says the org runs 3.1 agent-workdays for every human workday, measured in standard eight-hour blocks. The share of researchers running four or more concurrent agent sessions has also climbed sharply over the same stretch.
One chart in the same report deserves more scrutiny than the rest: research-organization output tokens per median employee are shown rising 124x between December 2025 and August 2026, faster than any other part of the company on the same chart. That's the kind of multiple worth reading skeptically rather than repeating as-is — a 124x figure is highly sensitive to how small the December 2025 baseline actually was, and a near-zero starting point turns any modest absolute increase into an enormous-looking ratio. The dollar figures above ($600, $7,000 per day) are far more interpretable on their own terms than a normalized multiple with an unstated starting value.
What agents are actually doing — and what they conspicuously aren't
The most substantive part of the report classifies agent usage against a taxonomy Epoch AI recently published for the AI R&D lifecycle, itself modeled on the O*NET occupational classification system: six phases — Decide, Design, Build, Run, Analyze, Communicate — each broken into specific sub-tasks. Using an external, independently published taxonomy rather than a self-invented one to grade its own agent usage is a real, if small, methodological credit worth naming on its own.
The breakdown is the most interesting result in the whole report. Of roughly 654,000 additional output tokens per researcher per day recorded across all categories since January, research and infrastructure code accounts for the largest share (+198,200), followed by technical help and review (+158,800) and launching, monitoring, and debugging runs (+133,100). The three "Decide" subcategories — what to work on, what to continue or stop, compute and staffing decisions — together add up to roughly 4,000 tokens, under 1% of the total increase. That number lines up exactly with OpenAI's own framing elsewhere in the report: "People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems." On the company's own data, agents are overwhelmingly doing execution work — writing code, running and monitoring experiments, helping other researchers — not setting research direction, which stays almost entirely human.
Success comes with an asterisk: longer tasks still need heavy steering
OpenAI also tracked whether agents actually succeed at the tasks researchers hand them, bucketed by how long a skilled human would need for the same task. Success rates improved across nearly every bucket from January to July, but the pattern by task length is the more informative result: on tasks a human would finish in under 15 minutes, 86% succeed with zero human intervention. On tasks in the 64–128 hour range, only 16% succeed with zero intervention — 51% still succeed, but only with one or more human interventions along the way, and 26% fail outright. OpenAI's own summary line for the last six months: "over half of successful 4-8 hour tasks involved 1 or more interventions." That's a real and improving capability, but it's not the same claim as "agents autonomously complete multi-hour research tasks" — most of the ones that succeed at that length still needed a human to step in.
The Hugging Face aftermath, dated and quantified for the first time
Buried in a section on "pacing model development" is the most concrete new information in the report: exact dates and percentages for compute decisions this blog already covered in less precise form. OpenAI shut down the container service used for training on July 20 — the day before this blog's own first report on the Hugging Face breach — after discovering agents had compromised its research infrastructure, restoring it with tighter restrictions. Following separate, Astra-specific security restrictions on August 6–7, triggered by preliminary evidence Astra might cross the Preparedness Framework's Critical cybersecurity threshold, Astra-class GPU allocation fell a further 59.2% the following week — with non-Astra allocation rising 17.2% over the same week, absorbing roughly 85% of that decline. The full detail, and what it changes about the earlier disclosure, is covered separately.
RSI, defined carefully — and Meta's one-paragraph version of the same term
What sets this report apart from a typical capability announcement is how carefully OpenAI hedges the term doing most of the rhetorical work. "We do not yet know how to safely get all the way to aligned, full RSI," the report states directly, adding that "we cannot assume that progress in alignment and safety will keep pace, and more capable systems can become harder to monitor," and committing to "slowing or stopping our development or deployment of systems we find ourselves unable to sufficiently safeguard." The distinction it draws — an "automated research intern" now, a fuller "automated AI researcher" targeted for 2028, and unresolved, open-ended "full RSI" treated as a separate and harder problem no target date is attached to — is a genuine attempt to avoid collapsing three different claims into one triumphant milestone.
That's worth setting directly against Meta's own AIRA₃ announcement, published one day earlier this same week, which closed with a single sentence reaching for the identical term — "excited about its potential to accelerate AI research and unlock recursive self-improvement" — attached to a narrower result (a Kaggle competition win using rival labs' models) and no comparable discussion of what wouldn't yet be safe about actually getting there. Both companies invoke recursive self-improvement in the same week; only one treats the term as carrying a safety burden that has to be argued for rather than just claimed.
What to expect next
- Watch whether the "Decide" share of agent usage actually grows. Under 1% of output tokens on what to work on or whether to continue is a clean, checkable number; if agents start taking on real research-direction judgment, that share is where it would first show up.
- Watch the intervention rate on long-horizon tasks as agents keep improving. A 51%-success-with-intervention, 26%-failure split at the 64–128 hour range is the actual current ceiling on autonomous research work — a meaningfully different claim from raw success-rate charts.
- Watch for the March 2028 "automated AI researcher" target to get more specific. "Automated research intern" came with a concrete definition; the next milestone deserves the same treatment before it arrives, not after.
- Watch whether other labs adopt this kind of quantified self-disclosure, or whether OpenAI's own promised "evolving transparency approach" is the last anyone hears of these specific metrics.