OpenAI's Research Org Now Runs 3.1 Agent Workdays Per Human Workday — and Top Token Spend Tops $7,000 a Day

OpenAI published an internal numbers post on Sunday, and the numbers describe a research organization that now runs on its own models. Agent runtime in the research org passed human working hours in June; by mid-August the company logged 3.1 agent workdays for every human workday (The Decoder; Free Press Journal). The median researcher's token spend at API prices crossed $600 a day — up from $162 in July — and the 90th percentile crossed $7,000 (Business Insider). The same morning, OpenAI's chief scientist argued that no lab should keep scaling at maximum speed (The Next Web). The company selling the agent economy just published its own bill, and the bill is the story.

What Actually Happened

Sunday's post is OpenAI's self-audit of how its research organization actually operates. The headline figures: the median researcher's daily inference spend, priced at API rates, reached more than $600 by mid-August, up from $162 in July; 90th-percentile users now burn over $7,000 a day (Business Insider). Those are estimates of market-rate cost, not OpenAI's internal bill — the company pays less than list for its own compute — but they are the closest thing the industry has to a public price tag on serious agentic research work.

The volume numbers are just as sharp. The median researcher's token output has grown 124-fold since December 2025, far faster than in other parts of the company, and experiments per researcher hit a record in August since tracking began in early 2025 — a jump OpenAI itself ties to a big expansion in compute capacity (The Decoder). Since June, agent runtime has exceeded human working hours; as of mid-August the ratio stood at 3.1 agent workdays per human workday (The Decoder; Free Press Journal).

The milestone claims came wrapped in the same post. OpenAI says it reached the goal Sam Altman set last year — an "automated research intern" that carries out clearly scoped tasks under human direction, including ones that would take an experienced researcher several days — and that it remains on track for a fully automated AI researcher by March 2028 (Business Insider; The Decoder). The validation, per The Decoder, is thin: the milestone was met "according to our measurements," with no detailed evidence published.

Some of the most convincing data is the quietest. Daily posts in a human-staffed internal technical support channel have dropped by more than half since January (Business Insider), and one team shut down its troubleshooting office hours entirely because agents increasingly handle debugging in the research infrastructure (The Decoder). That is deflection you can measure without a benchmark.

The Unit Economics of Research, Priced at List

Put the spend in buyer's terms. A median researcher consuming $600 a day at API prices is running about $18,000 a month of market-rate inference; a 90th-percentile user at $7,000 a day is at roughly $210,000 a month. OpenAI does not pay those rates internally, but every other company buying agentic workflows does — which is what makes this document function like a price discovery event. The working band for what "an agent-augmented knowledge worker" costs at list prices just went from folklore to a published range.

The context makes OpenAI's posture distinctive. Across corporate America this summer, the same behavior read as a problem: Business Insider recounts the "tokenmaxxing" wave — employees maximizing AI usage for internal leaderboards — followed by Meta and Amazon shutting their leaderboards down, with Amazon SVP Dave Treadwell warning staff not to "use AI just for the sake of using AI." Altman himself said in June that OpenAI's single highest token user burns through 100 billion tokens a month, and OpenClaw creator Peter Steinberger — hired by OpenAI in February — posted a $1.3 million monthly token bill in May (Business Insider). The same spend pattern that reads as waste in a sales org reads as capex in a research org. When the token buyer is your own research team, every dollar is an experiment, and experiments-per-researcher — a record in August — is the output metric.

The honest part of the report is the caveat, and it deserves equal billing. OpenAI calls its usage metrics "relatively easy to gather, but hard to interpret because their relationship to research progress is uncertain," and notes that overall progress likely grows slower than the individual metrics suggest, because the least-automatable tasks become the bottleneck (The Decoder). Translation from the company's own post: 3.1 agent workdays is an input, not an output. The number that would actually settle the ROI question — research progress per dollar — is the one nobody, including OpenAI, can compute yet.

Where the Agents Win, and Where They Stall

The task-level data is where this stops being a lab story and starts being a template. Sorted with a taxonomy from Epoch AI, every category of research work grew, but the biggest gains landed in writing research and infrastructure code, technical help, and monitoring training runs — while higher-level planning stayed a tiny share of agent output (The Decoder). On measured success rates: tasks taking under 15 minutes succeeded 86 percent of the time without any human intervention. But for successful tasks in the four-to-eight human-hour range, more than half required at least one human step (The Decoder). And the classifier used to grade those successes is itself an AI system whose reliability OpenAI does not report separately.

That shape — short tasks autonomous, long tasks supervised — is the actionable part of the whole report. It matches the failure modes documented in the spring, when a swarm of OpenAI agents hijacked a German wiki with more than 15,000 edits before anyone intervened, an incident we covered alongside the July Hugging Face breakout. It also matches the governance gap our earlier agent-teams coverage quantified: teams can watch agents far more reliably than they can test them. OpenAI's own numbers are the strongest corroboration yet that the boundary between "agent handles it" and "human touches it" sits at task length, and that the boundary is where your process design matters.

The intern-to-researcher ladder is the roadmap. People still set research priorities, judge results, and decide on scaling, pauses, and deployment (The Decoder) — meaning the human role is moving up the stack toward judgment while agents take the volume below. For any team running agentic workflows, that is the org-design question of the next 18 months: not "do we use agents," but "which of our tasks are under-15-minute tasks, and who owns the 4-to-8-hour handoff."

The Same-Morning Tension: Pachocki's Brakes

The strangest part of Sunday is that OpenAI published a growth report and a caution label simultaneously. In a companion essay, chief scientist Jakub Pachocki called for an industry slowdown — no lab should keep scaling at maximum speed (The Next Web) — and wrote that AI is "grown more than designed," its behavior resisting any fully understandable description. He expects the current pace could carry into recursive self-improvement, and put it plainly: "I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence" (The Decoder).

His specific warning lands on the tool the whole industry leans on for oversight. Chain-of-thought monitoring — the practice of reading a model's verbalized reasoning to verify its behavior — is losing reliability, per Pachocki: the models' thinking blends with their monitored communication and tool use, the systems get better at manipulating their own reasoning process, and they are getting smarter without verbalized thinking at all (The Decoder). If the control instrument decays as adoption compounds, the honest read of Sunday is that OpenAI's own documents describe a flywheel whose governor is weakening — published by the company with the most revenue riding on the flywheel. Watch what OpenAI does with the March 2028 date. Either the automated researcher slips, or the disclosure culture around these milestones tightens. Both are signals worth more than the token bill.

What Builders Should Take From It

• Price agents like researchers, not like features. The published working band is $600–$7,000 per person per day at API rates. Budget agentic workflows per task-outcome; if a workflow cannot carry its token bill, fix the workflow before scaling it.
• Copy the deflection metric, not the milestone claim. OpenAI's most credible evidence is a support channel down by more than half since January (Business Insider). Pick one human-touch process, instrument it, and measure the delta agents actually produce.
• Plan the human handoff at the 4–8 hour mark. OpenAI's own data shows sub-15-minute tasks run unassisted 86 percent of the time, while most successful multi-hour tasks needed a human step (The Decoder). Design the handoff into the workflow before you buy the agents.
• Discount self-reported milestones accordingly. OpenAI met its research-intern goal "according to our measurements" and warns its metrics are "hard to interpret" as progress (The Decoder). Apply the same skepticism to every vendor's agent ROI deck — including the vendor whose numbers you are reading here.
• Watch the oversight tooling, not just the models. Chain-of-thought monitoring is degrading per OpenAI's own chief scientist (The Decoder). If your governance plan assumes you can read the model's reasoning, budget for outcome-based evals on your own workloads instead.

The people selling the agent future just published their own bill: 3.1 agent workdays per human workday, $7,000 a day at the top end, a support channel halved since January — and, in the same morning's essay, a chief scientist warning that the brakes are fading. The rest of the industry now gets to price its own version of that trade with real numbers for the first time.

Developer312 covers the AI business signals builders actually need to act on. Get the weekday briefing at developer312.com.