Research

OpenAI researchers say they'd give up 40% of colleagues before their coding agents

OpenAI's second research-acceleration report shows median agent token use at $2,400/day, 46% autonomous success on 2-4 hour tasks — and researchers who'd trade 40% of colleagues to keep the tools.

OpenAI's own researchers now say they value their AI tools more than nearly half their colleagues
OpenAI's own researchers now say they value their AI tools more than nearly half their colleaguesAI-generated
By Sophie Lindqvist5 min read

Why it matters

  • Median OpenAI researcher's theoretical token consumption reached $2,400/day in October, up 62% month-over-month; top 10% hit $23,000/day.
  • Autonomous agent success on 2-4 hour tasks rose from 23% to 46% between March and October; 4-8 hour tasks rose from 13% to 35%.
  • In a survey of 101 employees, researchers said they'd give up 40% of team staff or 30% of compute rather than lose coding agents for three months.
  • Key conceptual ideas for GPT-6 Astra and GPT-6.2 Astra came exclusively from humans; agents succeed 79% at training monitoring but only 25-29% at complex engineering.
  • METR estimated AI use nearly doubled development pace at Anthropic after an eleven-week internal study.

OpenAI's own researchers would rather give up 40 percent of their team's staff — or 30 percent of their compute capacity — than work three months without their current coding agents. That's the headline finding from OpenAI's second report on internal research acceleration, published with the company's own usage data and survey results. The numbers describe a research organization that has quietly restructured itself around AI agents, while still hitting hard limits on what those agents can actually do.

The stakes here go beyond one company's internals. OpenAI says it published the data to document early indicators of so-called recursive self-improvement (RSI) — the scenario where AI systems meaningfully speed up the development of future AI systems. If that loop is starting to close, OpenAI's researchers are among the first people it would show up in. The report is, in effect, a self-audit of that loop.

How much agent work is actually running?

Usage of OpenAI's internal tools — chiefly the coding agents Codex and Dots — has surged in recent months. At regular API list prices, the median researcher would have consumed $2,400 worth of tokens per day in October, up 62 percent from September. Among the top 10 percent of heaviest users, that theoretical daily value doubled to $23,000.

Converted into working hours, the agents reportedly run the equivalent of 5.7 eight-hour workdays per researcher per calendar day. In the pre-training and post-training teams specifically, the figure rises to about eight workdays per researcher per day. More than two-thirds of tracked researchers ran at least four agents in parallel at peak usage.

The token figure is theoretical — OpenAI is pricing its own internal consumption at list rates, not paying an external vendor. Still, the comparison the company draws is striking: in September, the theoretical API equivalent of agent runtime was twice as high as researchers' regular salary costs. Two caveats temper that. The calculation excludes equity packages, which make up 70 to 80 percent of total compensation at OpenAI. And measured against the company's total research compute, internal agent use accounted for just 2 percent.

Takeaway: agents already out-work researchers by hours and out-cost them by list-price tokens, but they remain a rounding error in OpenAI's overall compute budget.

What can the agents actually do?

Between March and October, the estimated success rate of autonomous agents working without human intervention rose sharply. On tasks taking two to four hours, success climbed from 23 percent to 46 percent. On four-to-eight-hour tasks, it went from 13 percent to 35 percent.

A breakdown using an Epoch AI taxonomy shows the catch: the systems mainly succeed at simple tasks. Agents complete training-run monitoring 79 percent of the time and answer technical questions independently 72 percent of the time. The numbers drop fast once the work gets harder:

  • Building evaluation and reinforcement-learning environments: 29 percent success
  • Writing code for inference and serving systems: 25 percent success

So the doubling of headline success rates reflects real progress, but it's progress concentrated in routine monitoring and question-answering, not in the engineering work that's hardest to automate.

Do researchers actually prefer the agents to people?

The report includes two small, non-representative internal surveys that try to answer this. In the larger one, 101 employees were asked what they'd sacrifice to keep their coding agents. At the median, researchers said they'd give up 40 percent of their team's staff or 30 percent of their compute capacity rather than go three months without the agents.

A separate quick poll of 28 employees posed a starker choice: the latest internal model or a mid-career human engineer. Twenty-seven of 28 picked the model.

Both surveys are small and OpenAI flags them as non-representative. But they point in the same direction as the usage data: the tools have become load-bearing, and the people using them know it.

Where the agents fail — and why humans stay in the loop

Qualitative interviews with ten researchers laid out the downside of this dependence. The models, the interviewees said, lack any "taste" for good research ideas and for code and data quality. In practice that means agents get lost in unnecessary steps, make wrong assumptions in analyses, and submit subpar "AI slop" code changes along with unusable data artifacts. Close human oversight is essential, the researchers said, to avoid wasting compute cycles.

The clearest evidence of the gap is in how OpenAI's newest models were actually conceived. The key conceptual ideas for GPT-6 Astra and GPT-6.2 Astra came exclusively from humans in the post-training and safety areas, according to the teams interviewed. Only in pretraining for GPT-6.2 Astra did individual researchers say they gained important insights through dialogue with Codex.

Takeaway: agents have absorbed much of the labor, but the judgment about what's worth building still sits with people — at least for now.

How this fits the wider picture

OpenAI cautions in its conclusion that the data gathered so far on human-AI collaboration may not extrapolate linearly. The company points to recent rapid AI breakthroughs in mathematics and considers it possible that similarly abrupt advances could occur in deep-learning research itself as models gain more autonomy. That's a carefully hedged way of saying the current 25-to-29-percent failure zones might not stay hard forever.

There's external corroboration that AI is already compressing research timelines. The evaluation organization METR, after an eleven-week look inside competitor Anthropic, recently estimated that AI use there had nearly doubled the pace of development. And OpenAI's first research-acceleration report, published in early September, declared that the company had reached its self-imposed goal of an automated AI research intern.

That first report came with a warning attached. Chief scientist Jakub Pachocki said in an accompanying post that OpenAI risked losing control, after an audit revealed numerous gaps in the oversight system. The second report's findings on "AI slop" and the need for constant human supervision suggest those gaps haven't fully closed — even as the organization bets more of its daily workload on the agents.

Original: openai.com

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

778 articles

Related articles

  1. OpenAI Says Agents Have Replaced Chatbots as Its Default Work Tool
  2. OpenAI Predicts AI-Made Discoveries by 2026 as Intelligence Costs Plunge
  3. OpenAI Study: Researchers Say ChatGPT Cuts Their Work Time Nearly in Half
  4. OpenAI's Enterprise Report: 1 Million Business Customers, 320x API Growth
  5. OpenAI Launches GDPval, a Benchmark Built From Real Jobs

« Previous article