Why OpenAI’s $23,000 Daily AI Token Burn Could Signal the Next Billion-Dollar Tech Goldrush—or a Massive Blowup
Ever wonder what it feels like to burn through $23,000 in tokens every single day just to get your coding done? That’s exactly the reality for OpenAI’s top researchers, according to a recent internal report that reveals a jaw-dropping 105% spike in their daily token consumption since September. Talk about putting your money where your code is! These guys aren’t just tinkering around—they’re diving headfirst into what the report calls “Research acceleration: The jagged frontier,” showcasing how deeply AI-driven coding agents have woven themselves into the lab’s DNA. What’s astounding is not just the sheer volume but the speed at which these agents are taking over tasks that once took hours of human effort—raising an eyebrow at how we might soon need to rethink the whole work-team equation. So, is this the future of research productivity, or are we just getting started on a spending spree fueled by machine brains? Hold onto that thought because this AI revolution is accelerating faster than most dared to imagine. LEARN MORE

An internal OpenAI report shows heavy-use researchers more than doubled their coding agent consumption since September
The most agent-hungry researchers at OpenAI now run through about $23,000 worth of tokens per day. That figure comes from an internal report dated November 18, 2026, and it is up 105% from September.
The report, titled “Research acceleration: The jagged frontier,” tracks how deeply coding agents have embedded themselves in OpenAI’s own research staff.
The numbers behind the surge
The headline metric covers the 90th percentile of OpenAI researchers. These are the heaviest users, but not outliers at the extreme edge.
Their average daily consumption hit $23,000 in tokens, valued at API list prices. Back in September, the same group was at $11,000 per day.
The “API list prices” qualifier matters. The dollar amounts reflect what this usage would cost an outside customer paying the public rate.
The middle of the pack is moving too, just less dramatically. Median researcher usage climbed to $2,400 per day, a 62% jump from $1,500 in September.
Perhaps the most striking figure is about parallelism. The report puts the average at 5.7 agent workdays per researcher per calendar day.
What the agents are actually doing
According to the findings, agents are now completing almost half of research tasks that would typically take a human two to four hours.
The researchers themselves appear to value the tools enormously. According to the report, staff said they would give up 30% of their team’s compute, or 40% of their team’s headcount, to keep access to frontier agents for three months.
A trend that has been building all year
This is not OpenAI’s first look at the phenomenon. A September 2026 report found that output tokens grew 124-fold between November 2025 and August 2026.
The pattern also tells a story about adoption. Growth at the 90th percentile outpaced growth at the median by a wide margin, 105% versus 62%.
What this means for the AI industry
The most immediate implication is about how AI labs build AI. OpenAI is using its own agents to accelerate its own research, and the internal usage curve is steep.
The headcount trade-off finding deserves attention from anyone watching the labor side of AI. When researchers say they would accept 40% fewer teammates to keep their agents, it reframes how organizations might think about staffing versus tooling.
That does not mean OpenAI plans to cut staff. The report frames it as a hypothetical preference, and nothing in the findings indicates a hiring change.
They have privileged access to the newest models, deep expertise in prompting them, and no meaningful cost barrier to running them. The average enterprise developer paying list price faces a very different calculation.
If $23,000 a day at list prices is what peak productivity looks like, the economics for outside customers depend heavily on how token prices move from here.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.




Post Comment