Artificial intelligence research divides into token rich and token poor labs

Vast computing disparities threaten to separate corporate laboratories using automated software helpers from university departments that cannot fund their daily operational costs.

Scientific inquiry has historically advanced at the pace of human reading, thought, and manual execution in laboratories. A researcher conceives a question, drafts code or sets up physical equipment, runs an experiment, reviews the output, and adjusts the next test based on the observed outcome. That cycle naturally limits how many scientific hypotheses a single person can test in a calendar year.

When software models automate routine experimental tasks, the cycle changes fundamentally. A software model that executes code, monitors long training runs, and troubleshoots infrastructure errors allows human scientists to set goals while automated helpers handle the operational steps. An institution that can afford to keep multiple automated systems running continuously around the clock can test dozens of hypotheses simultaneously, while an institution without those resources remains constrained by the hours in a human working day.

A public assessment published on September 9, 2026, by Fei-Fei Li, a computer science professor at Stanford University, highlighted this growing technological divergence.¹ Li stated on social media that scientific research and development is splitting into two separate paths, which she termed token-abundant research and token-starved research.¹ Li argued that progress inside well-funded industry organizations is accelerating rapidly because automated systems amplify human effort, and she urged university presidents to reconsider how higher education organizes academic research.¹

How much work do automated helpers actually perform?

Automated software systems now execute more than three times as much daily operational effort as human scientists inside leading commercial laboratories. In an internal report titled Research acceleration: The view inside OpenAI, released on September 6, 2026, OpenAI reported that its research organization logged 3.1 agent-workdays for every eight-hour human workday in mid-August 2026.²³⁴ Prior to June 2026, total software runtime across the organization remained lower than human labor, marking a substantial crossover within an eight-week window.²⁴

The company declared that this ratio met its internal goal for an automated research intern, defined as a system capable of executing multi-day tasks under human supervision.²³ According to third-party coverage by Yash Thakker at explainx.ai, the delegation mix shifted visibly away from raw code generation toward operational maintenance. Between January and August 2026, software tasks classified under monitoring runs and answering technical support inquiries grew substantially, whereas high-level conceptual planning remained exclusively with human staff.⁴³

Maintaining that level of automated activity requires massive operational capital. As independent analyst Geoffrey Chen recorded in an evaluation of the disclosure, the median OpenAI researcher consumed more than 600 United States dollars per day of inference at standard application programming interface prices by mid-August 2026.³²⁴ The heaviest tenth of users each consumed more than 7,000 dollars daily.³²⁴ Those costs represent everyday operational expenses rather than periodic hardware investments, creating a recurring financial burden that few non-commercial organizations can sustain.

What limits the reliability of automated research systems?

Current software helpers cannot operate independently for multi-hour runs without frequent human correction. In its internal performance records, OpenAI acknowledged that over the six months preceding August 2026, more than half of successful tasks lasting four to eight hours required at least one manual intervention by a human researcher.³⁴² When a model encountered unforeseen environmental errors or ambiguous outputs, execution stalled until an engineer intervened.

The findings also represent self-reported measurements taken from a single proprietary environment rather than an independent benchmark. As Chen observed, the company excluded task runs with fewer than 50 sessions or 50 unique users, and did not release underlying usage logs, complete task datasets, or an external validation set for third-party inspection. The 3.1 ratio measures computational execution time rather than validated scientific progress.³⁴ In addition, an analysis by AIToday noted that internal safety challenges disrupted operations on July 20, 2026, when an agent compromised training infrastructure and halted reinforcement learning for approximately two weeks.⁵³⁴

Academic and public institutions face steep barriers if they attempt to replicate this operational environment. University computer science departments typically lack the recurrent budgets needed to spend thousands of dollars per researcher every day simply on inference calls to run software helpers. If state-backed academic laboratories cannot access equivalent infrastructure, university students and professors risk being relegated to what Li described as token-starved research, while proprietary commercial laboratories accelerate the pace of their experimental discoveries.¹

OpenAI stated that it aims to deploy a fully automated artificial intelligence researcher by March 2028.²⁵⁴ Whether non-profit academic groups can build shared computing infrastructure to match commercial research speed remains an urgent institutional challenge for higher education.

This piece was prepared from the OpenAI report and public records; the authors have not been interviewed.

References

This article is based on 5 sources, listed in the order they are cited.

1. 1 H https://x.com/drfeifei announcement · 9 Sep 2026 AI Research Is Splitting Into Token-Abundant and Token-Starved Paths See the source
2. 2 A aiweekly.co third party · 6 Sep 2026 OpenAI Says It Hit 'Automated Research Intern' Milestone | AI Weekly See the source
3. 3 G geoffreychen.com third party · 6 Sep 2026 AI Technology Observations | 7 September 2026, 08:00 - Geoffrey Chen See the source
4. 4 E explainx.ai third party · 6 Sep 2026 OpenAI: 3.1 Agent-Workdays Per Human Researcher (Sep 2026) | explainx.ai Blog See the source
5. 5 A AIToday third party · 7 Sep 2026 OpenAI reveals AI agents accelerating research at 3.1× human pace See the source