OpenAI published a self-report on September 6 called “Research acceleration: the view inside OpenAI,” and buried in the productivity-milestone framing is a genuinely revealing admission: the same model that crossed OpenAI’s own Critical cybersecurity threshold three days after this blog covered it is now expensive enough to run safely that it’s crowding out OpenAI’s own training budget.
The headline number, and the caveat that matters more#
As of mid-August, OpenAI’s research organization is running 3.1 agent-workdays of effort for every workday of human labor, measured against a standard 8-hour day. The median researcher is spending more than $600 a day on inference at API prices; the 90th percentile researcher is burning through more than $7,000 a day. OpenAI frames this as hitting a goal it set last fall — fielding an “automated research intern” capable of executing well-defined, multi-day research tasks under human direction, with a stated target of an actual automated AI researcher by March 2028.
The caveat OpenAI attaches to its own number is the part worth reading twice: more than half of successful tasks in the four-to-eight-hour range still involved at least one human intervention. The system, in OpenAI’s own words, “does not need to originate a research program, decide which unanswered questions deserve resources” — it executes, it doesn’t direct. Read the 3.1x figure as a measure of activity, not autonomy, and OpenAI’s framing holds up. Read it as “AI is now doing three times the research work of humans” — the version that’ll circulate on social media — and it doesn’t.
The part OpenAI didn’t put in the headline#
Here’s the more interesting disclosure. On August 7, preliminary evidence that GPT-6 Astra might cross the Critical cybersecurity threshold under OpenAI’s Preparedness Framework triggered additional model-specific security restrictions, forcing Astra to run only in higher-security research environments. In the week that followed, Astra-class GPU allocation for the reinforcement-learning workloads OpenAI analyzed fell 59.2%. Allocation to other model classes rose 17.2% to partially absorb the shift, offsetting roughly 85% of the decline — but total capacity in that workload still took a real, measurable hit.
That’s not a hypothetical safety-tax thought experiment. It’s OpenAI’s own numbers showing that classifying a model as Critical-risk has an immediate, quantifiable operational cost even inside the lab that built it — computing capacity had to be physically reallocated away from the model that scored 100% on ExploitBench, weeks before that model shipped to every paying ChatGPT tier with little more than an admin toggle and a refusal prompt standing between it and general availability. The internal handling (throttle GPU access, force higher-security environments) and the external handling (ship broadly, gate the advanced capability behind an alpha program) tell two different stories about how seriously the same classification is being taken, and only one of those stories is visible to customers.
“We do not yet know how to safely get all the way there”#
The report’s most quotable line is also its most honest: OpenAI states plainly that it “does not know how to safely get all the way to aligned, full” recursive self-improvement, and that research progress “will not necessarily scale with the raw activity metrics” it just spent the whole post promoting. That’s a real disclosure, not boilerplate — a lab publishing its own acceleration numbers in one breath and admitting it doesn’t have a safe path to the destination those numbers are supposedly accelerating toward, in the next.
Why this matters beyond OpenAI’s research org#
Every one of these dynamics — cost concentration among heavy users, GPU/compute contention between capability tiers, the gap between “activity” and “autonomy,” and the operational overhead of taking a Critical-risk classification seriously — maps directly onto how any team running Claude Code, Codex, or a comparable agent at scale should be thinking about their own agentic workflows. The 90th-percentile-to-median cost spread (12x) inside OpenAI’s own research org is a preview of what happens to any organization’s inference budget once agentic coding shifts from occasional assistance to concurrent, always-on sessions. And the Astra GPU-allocation story is a rare, numbers-backed answer to a question labs usually only discuss in the abstract: what does actually respecting a Critical-risk classification cost, in resources you can count? Anthropic’s answer to the same underlying tension — restricting Mythos-class capability to a gated research population rather than shipping it broadly with a refusal prompt — looks more and more like the more expensive but more defensible choice, and now there’s a real number attached to what “expensive” means.
Sources: Research acceleration: the view inside OpenAI; OpenAI Hits Goal of Building an “Automated Research Intern” — Unite.AI; OpenAI just hit a milestone on the road to self-improving AI — Help Net Security; OpenAI Says It Has Reached an Automated Research Intern — Data Studios; OpenAI Says Coding Agents Now Do 3.1 Workdays for Every Human Workday — Inside AI News; this blog’s own prior coverage of GPT-6 Astra’s Critical cybersecurity classification.