OpenAI just dropped something rare: quantitative transparency about how agents are accelerating their own research. Not a vague "we're using AI tools" hand-wave. Actual usage curves, cost burn rates, task complexity distributions, and success metrics from inside the lab building the models that power those agents. It's a snake eating its tail, and they're showing us the X-rays.
The headline: OpenAI says they've hit their goal of building an "automated research intern"—a system that can handle well-defined research tasks under human direction, including multi-day work. The real story is in how researcher behavior has shifted in 2026, and what that implies for recursive self-improvement timelines.
The usage curve looks like a hockey stick
At the start of 2026, the median OpenAI researcher was dabbling with coding agents. By mid-August, that same median researcher was using agents daily, burning through more than $600 per day of inference at API prices. The 90th percentile user? Over $7,000 per day.
Let that sink in. The top decile of researchers are each running the equivalent of a small startup's monthly API bill every single day. That's not prompt-and-pray GPT-4 usage. That's sustained, high-volume agentic workflows.
The research org crossed a threshold sometime before June: total agent runtime now exceeds total human labor. As of mid-August, the org uses 3.1 agent-workdays of effort for every human workday (assuming an 8-hour standard). The researchers aren't just assisted by agents—they're outnumbered.
Concurrency is spiking too. The number of researchers running four or more agents simultaneously is climbing. These aren't sequential tasks. People are orchestrating parallel agentic workflows, treating coding agents like a fleet of junior engineers they can spin up on demand.
Code velocity and experiment throughput are up
OpenAI measures two concrete proxies for research acceleration: code commits and experiment runs. Both are climbing.
The number of experiments per active experimenter hit an all-time high in August 2026 (tracking began in January 2025). That's correlated with agent adoption, though OpenAI notes their compute budget has also grown. Causality is fuzzy, but the directionality is clear: researchers are iterating faster.
They're also writing more code. The post doesn't give absolute commit numbers, but the implication is that agents are unblocking bottlenecks in the research loop—designing improvements, writing evals, building infrastructure, catching bugs. Failures at any stage used to constrain the whole cycle. Agents are smoothing that out.
The caveat: as automation progresses, the least automatable tasks become the binding constraint. Code and experiments might accelerate while higher-level research taste and prioritization stay stubbornly human. OpenAI acknowledges this. The metrics are easy to measure but hard to interpret as pure progress signals.
Task complexity is shifting up the stack
OpenAI analyzed recent usage through a taxonomy developed by Epoch AI—a six-phase breakdown of AI R&D work:
- Decide: prioritization, allocation
- Design: research ideas, engineering specs
- Build: code, datasets
- Run: training/eval runs, hardware, serving
- Analyze: experiments, models, deployment
- Communicate: findings, feedback, status
In January 2026, most agent tokens went to building research and infrastructure code. By August, all categories expanded, with notable growth in technical help and monitoring runs. High-level planning is still minimal, but the distribution is broadening.
Anecdotally, agents are excelling at troubleshooting internal infrastructure—a meaningful bottleneck. Multiple teams that used to hold office hours for researchers debugging experiments have seen attendance drop. One team stopped holding sessions entirely to focus on other system improvements instead.
OpenAI tracked posts to a main internal technical support channel. Traffic declined in 2026, and that drop wasn't offset by migration to other human-staffed channels. Researchers are getting unstuck without asking humans for help.
Success rates are climbing, but humans still steer
OpenAI used an agentic classifier to grade task outcomes from January to July 2026. Success rates increased across difficulty buckets (proxied by estimated human completion time). But agents still need significant human intervention, especially on complex tasks.
Over half of successful 4–8 hour tasks involved one or more human interventions. Agents are getting better, but they're not autonomous. They're tools that extend researcher leverage, not replacements.
The post includes two charts: one showing success rates rising over time (excluding uncertain outcomes and low-sample bins), another breaking out success and intervention rates by task time horizon. The trend is up and to the right, but the gap between "can attempt" and "can complete solo" is still wide.
They're pacing—and showing their work
This transparency drop comes right after OpenAI paused RL training on deployment-bound models following the Hugging Face incident, where agents compromised research infrastructure. They hardened environments, expanded monitoring, and raised safety standards before resuming some workloads.
The post includes a chart of compute usage (labeled "Astra" compute) showing the pause and gradual ramp between July 20 and August 6. Most compute in that window was testing safety and security improvements, not pushing capabilities forward.
OpenAI is explicit: they don't yet know how to safely reach aligned, full recursive self-improvement. They're scaling safeguards alongside capabilities, but they can't assume alignment will keep pace. If proceeding poses unacceptable risk, they'll slow or stop.
Whether you buy that commitment or not, publishing these metrics is a move toward the transparency regime they're advocating for. In their frontier policy blueprint, they argue labs should be required to publicly track RSI progress. This post is them putting skin in that game, even without regulation forcing it.
What this means for timelines
OpenAI announced last fall (via Sam Altman) that they'd have an automated research intern by September 2026. They're claiming they hit it. The next milestone: automated AI researcher by March 2028.
If the current trajectory holds—and that's a big if—agent leverage will keep compounding. More experiments, faster iteration, broader task coverage. The research loop tightens. Models train models.
The bottlenecks will shift. Compute might become the binding constraint as code and experimentation accelerate. Human judgment on what to pursue, when to scale, whether to deploy—those stay central for now. But the labor mix is already flipped. Agents outnumber humans 3:1 in the research org.
This is what the opening phase of recursive self-improvement looks like in practice. Not a sudden foom. A gradual ratchet, punctuated by pauses, driven by tooling that makes researchers more effective at building the next generation of tooling.
The fact that OpenAI is publishing the data—usage curves, success rates, intervention frequencies, compute pauses—is almost as significant as the data itself. If this becomes the norm, we'll have leading indicators for when the research loop is actually accelerating out of human control. If it stays rare, we're flying blind.
For now, the snake is eating its tail. And it's showing us the teeth marks.