Skip to main content
OpenAI researchers analyze data on screens, reflecting on AI tools' value over human colleagues; a complex ethical dilemma.

Editorial illustration for OpenAI researchers value their AI tools more than half their colleagues

OpenAI's Coding Agents Outpace Human Researchers

OpenAI researchers value their AI tools more than half their colleagues

• 4 min read

OpenAI has put numbers on something its staff have apparently started saying out loud: the coding agents built inside the company now do more work, by some measures, than the humans running them. A second internal report on research acceleration, published by OpenAI itself, tracks how far tools like Codex and Dots have burrowed into the daily workflow of building new models. The company frames this as evidence toward recursive self-improvement, the idea that AI systems can speed up the creation of the AI systems that come after them.

The figures are blunt. Median researchers burned through $2,400 worth of tokens a day in October at standard API pricing, a 62 percent jump from September. Heavy users hit $23,000 a day.

Translate that into work hours and the agents are logging close to six eight-hour shifts per researcher per day, more in some teams. Most researchers tracked ran several agents side by side at peak times. What that usage says about how OpenAI's own staff now rank human versus machine contribution is where the report gets uncomfortable.

OpenAI's internal use of coding agents is growing fast, with theoretical token costs now exceeding researchers' base salaries. Even though teams are reluctant to work without these tools, the models still produce a lot of "AI slop" beyond routine tasks.

Why this matters

OpenAI putting a number on this is the interesting part. When your own researchers rate the agent above a chunk of their colleagues, and the token bill for that agent theoretically runs double what you pay those researchers, you've stopped talking about a coding assistant and started talking about a line item competing with headcount. That the comparison excludes equity, which is 70 to 80 percent of OpenAI pay, should make anyone citing this data pause before drawing budget conclusions.

It's a skewed yardstick dressed up as a clean metric. For developers and founders watching this, the lesson isn't "replace your team," it's that internal AI usage is getting expensive enough to track like payroll, even while the output still includes a fair amount of slop once you leave narrow, repetitive tasks. Treat this as OpenAI's own evidence that heavy agent use and real cost discipline haven't been reconciled yet, not as proof that agents are outperforming people.

The gap between theoretical API cost and actual value delivered is exactly where we'd want more scrutiny, not less.

Common Questions Answered

How do OpenAI researchers value their coding agents compared to their colleagues?

According to OpenAI's internal report, researchers now value their AI coding agents more than half of their colleagues, indicating that these tools have become integral to their workflow. The company has quantified this by showing that the theoretical token costs of these agents now exceed researchers' base salaries, demonstrating their significant computational and economic value to the organization.

What are Codex and Dots in the context of OpenAI's research acceleration?

Codex and Dots are coding agents that OpenAI has integrated into its daily workflow for building new models. These tools have become deeply embedded in researchers' processes, with teams now reluctant to work without them, representing a major shift in how the company approaches model development.

Why does OpenAI's comparison of AI agents to researcher salaries matter for budget decisions?

When the theoretical token costs of AI agents exceed researcher salaries and researchers value these agents above many colleagues, it transforms the discussion from a simple coding assistant to a direct line item competing with headcount for budget allocation. However, the comparison becomes more complex when considering that equity typically comprises 70 to 80 percent of OpenAI researcher compensation, which is excluded from this analysis.

What is recursive self-improvement as described in OpenAI's research acceleration report?

Recursive self-improvement refers to the concept that AI systems can speed up the creation of new models, essentially accelerating their own development cycle. OpenAI's internal data on coding agents demonstrates this principle in practice, showing how AI tools are directly contributing to faster model development and research advancement.

What limitations exist in the AI coding agents despite their high valuation at OpenAI?

Despite being valued highly by researchers and generating significant token costs, OpenAI's coding agents still produce considerable amounts of low-quality output or "AI slop" beyond routine tasks. This suggests that while these tools are valuable for productivity, they still require human oversight and refinement for non-standard work.

LIVE16:16Anthropic IPO expected to leave AI firm with USD 220 billion in cash