OpenAI Runs 3.1 Agents Per Researcher. What's Your Ratio?

OpenAI just told the world exactly how many AI agents it takes to replace the labor of one researcher. The number is 3.1. Ask your own team the same question and most operators go quiet.

That's not a knock — it's the actual gap. Companies have spent two years buying agent tools and approximately zero time building a way to measure what those agents produce. OpenAI, unusually, published its internal math [1]. It's worth taking apart, because the metric is more interesting than the milestone it's attached to, and because most of what it reveals travels straight down to a five-person team.

What OpenAI actually measured

The headline claim is that OpenAI hit its "automated research intern" goal — a system it can hand a bounded research problem that would take a skilled human a few days, and get usable work back without hand-holding every step [1]. OpenAI's chief scientist, Jakub Pachocki, had framed the target back in March: "a system that you can delegate tasks to that would take a person a few days" [2]. September 2026 was always the deadline. OpenAI says it made it.

But the milestone is the press release. The metric underneath it is the actual finding: OpenAI converted every hour of coding-agent runtime across its research organization into standard eight-hour workdays, then compared that total against total human labor hours for the same org. Before June 2026, combined agent runtime was still below combined human runtime — agents were a supplement, not a majority of the work. By mid-August, agents were logging more than three workdays for every one human workday [1].

That's not "we use AI a lot." That's a specific, auditable ratio, tracked the same way you'd track headcount utilization.

Where the 3.1x is actually coming from

OpenAI didn't just report one aggregate number — it broke research work into six phases: Decide, Design, Build, Run, Analyze, Communicate [1]. That taxonomy matters more than the topline ratio, because it tells you where agent labor concentrates and where it doesn't. Agents are eating the middle of the pipeline hardest — the execution-heavy phases where a well-specified task turns into code, into an experiment run, into a result. The phases that require judgment about what to try next are still much more human-weighted.

flowchart LR
  Decide[Decide<br/>what to try] --> Design[Design<br/>the experiment]
  Design --> Build[Build<br/>the code]
  Build --> Run[Run<br/>the experiment]
  Run --> Analyze[Analyze<br/>the results]
  Analyze --> Communicate[Communicate<br/>the finding]
  Communicate -.->|feeds back into| Decide

  classDef human fill:#fff,stroke:#000,stroke-width:2px
  classDef mixed fill:#fff,stroke:#000,stroke-width:2px,stroke-dasharray: 4 3
  classDef agent fill:#000,color:#fff,stroke:#000,stroke-width:2px

  class Decide human
  class Design mixed
  class Build agent
  class Run agent
  class Analyze mixed
  class Communicate human

This is the part most "we use AI now" claims skip. It's not that agents replaced researchers. It's that agent output concentrated in the most repeatable, most execution-bound phases, while the framing work — deciding what to work on at all — stayed a minimal fraction of what agents produced. That's a legible, defensible way to describe augmentation. Most companies can't produce this chart for their own AI usage, because they never instrumented the phases in the first place.

The spend curve nobody's job-titled yet

Here's the number that should worry finance more than the ratio does. By mid-August, the median researcher at OpenAI was spending more than $600 a day on coding-agent inference at API prices. The top 10% of users were burning through more than $7,000 a day [1].

That's not a SaaS line item. That's a spend curve shaped like a power law, sitting inside a company that can presumably absorb it. Most SMBs can't, and won't see a curve like that coming, because nobody's watching per-person inference spend as its own line — it's buried inside "software" until the invoice arrives and somebody asks what happened.

"Research intern" is a scoped claim — read what it doesn't say

Jakub Pachocki, OpenAI chief scientist, pictured in 2012. Photo: ICPCNews via Wikimedia Commons, CC BY 2.0. Story: OpenAI — Research acceleration: the view inside OpenAI.

Pachocki has been careful about what "research intern" means, and it's worth reading past the headline. It's not an autonomous scientist picking its own research agenda. It's a supervised system that takes a bounded objective from a human, works across code and experiments, and returns output for a human to grade [1][2]. The next milestone — a fully automated, multi-agent research system that can tackle problems "too large or complex for humans to cope with" — is scoped for around March 2028, not now [1][2][3].

A 3.1x agent-to-human ratio isn't proof you've automated research. It's proof you've built a very fast, very expensive intern that still needs a manager.

That distinction is the whole ballgame for anyone reading this as a buying signal. The ratio measures throughput on defined tasks, not judgment on undefined ones. Confusing the two is how a company ends up shipping an agent update nobody asked for, then walking it back a month later. I've watched the "over-automate, then pull back to draft-and-flag" cycle up close on client work — the fix is always the same: agents should escalate by default, not ship by default.

McKinsey's rebuttal: agent-hours is a vanity metric

Not everyone thinks "agent-workdays" is the number to chase. McKinsey has been arguing the opposite case: that token usage and agent-hours are leading indicators at best, and the metric that actually matters is cost per completed task — a number that keeps climbing across the workloads McKinsey tracks, because the appetite to throw agents at harder problems is basically insatiable [4]. Activity signals like agent-hours tell you adoption is happening. They don't tell you whether the business got anything back for it.

Both things are true at once, and OpenAI's own report is proof: 3.1 agent-workdays per human is a capacity signal — it tells you agents have become a majority input to the work. It says nothing about whether the output is 3.1x better, or just 3.1x more expensive to produce the same result. You need both numbers, not one instead of the other.

What this means for operators

You are not OpenAI. You don't have a research org burning $7,000 a day per person, and you shouldn't try to build OpenAI's dashboard. But the underlying discipline scales down cleanly, and most SMB teams I work with have none of it:

1. Pick one repeatable workflow — the kind of task pipeline that shows up in most agentic automation builds — and track agent-hours against the human-hours it used to take. You don't need six phases. Two or three will show you where the agent is actually doing the work versus where a human is still babysitting it.
2. Track cost per completed task, not cost per seat. A $200/month agent subscription that finishes four tasks a day is worse than a $400/month one that finishes twenty. Per-seat thinking hides that math; per-task thinking exposes it.
3. Watch your own top-10% users. OpenAI's spend curve had a long tail — most of the risk sits with your heaviest users, not your average one. If nobody's watching per-person inference spend, you'll find out the same way OpenAI's finance team probably did: from an invoice, not a dashboard. This is the same instinct that made a self-hosted n8n stack replace an $1,800/month Zapier bill — you don't manage what you don't measure, whether the line item is SaaS seats or token spend.

The teams that get the most out of agents in 2027 won't be the ones with the most subscriptions. They'll be the ones who can answer "what's our agent-to-human ratio, and what did it buy us" without having to guess.

Sources

[1] OpenAI — Research acceleration: the view inside OpenAI — openai.com [2] MIT Technology Review — OpenAI is throwing everything into building a fully automated researcher — technologyreview.com [3] TechCrunch — Sam Altman says OpenAI will have a 'legitimate AI researcher' by 2028 — techcrunch.com [4] McKinsey — Measuring agentic AI ROI beyond token costs — mckinsey.com

OpenAI didn't publish this number to help SMB operators build dashboards. It published it because a 3.1x ratio is a good headline on the way to a 2028 story about fully automated research. But the measurement habit underneath the headline is the actually useful export. If you're running agents in your business right now and you can't say what your ratio is or what it bought you, that's not a data problem — it's a decision to not look, and it's costing you the same way an un-audited SaaS stack does: quietly, then all at once.

The short version

• OpenAI's research org now logs 3.1 agent-workdays for every human workday, up from below 1.0 before June 2026
• The median researcher spends $600+/day on agent inference; the top 10% spend $7,000+/day
• Agents concentrate in the execution-heavy phases (build, run) — deciding what to work on is still a minimal fraction of agent output
• "Automated research intern" means bounded, supervised delegation — not an autonomous scientist; that's still scoped for around 2028
• McKinsey's counter-argument: agent-hours is a capacity signal, not a value signal — cost per completed task is what actually matters
• The exportable habit for smaller teams: track agent-hours against the human-hours a task used to take, price by task not by seat, and watch your heaviest users' spend before finance does