Hannah Xu, an investor who has spent recent weeks talking to researcher friends at frontier labs, argues that AI is already building the next AI inside those companies. She makes the case in an essay on X, “Inside the Frontier Labs: The First AGI Organizations,” posted on 30 October 2026. Her claim is that the labs have rebuilt their organization, their shared context and their infrastructure around agents, and that this is a different kind of company from anything the internet era produced.
Her test is blunt. Does what a company builds feed straight back into building its next version? An insurer that speeds up claims handling with AI still gains nothing toward a better AI, she writes. By her count, the frontier labs are nearly the only companies that qualify today.
The evidence she cites is a mix of public posts and lab figures. At OpenAI’s research organization, each human workday came with 3.1 agent workdays by mid-August. At Anthropic, Claude wrote over 80 percent of merged code by line count as of May. Zuxin Liu, an OpenAI researcher, said in September that he had not hand-written code for two months. He now tracks several agents and chooses which directions deserve effort.
The results are not uniformly strong. Xu points to an Anthropic experiment in which nine agents out-performed the human baseline on small models. The best idea then was carried to Sonnet 4.0 in a live production setup and gained 0.5 points, which sits within noise. She treats that as proof the loop is still fragile. An Anthropic employee told her the flip side: after a failure, you realize you cannot say what the machine did in the meantime.
Infrastructure is where the volume shows. Claude Code users approved 93 percent of permission prompts, Xu reports, which made each approval nearly meaningless, so Anthropic moved to sandboxes as the default plus a classifier that screens riskier actions. Merged changes per day climbed in May they numbered about 2,500, and by September above 30,000. On one pull request, the coding agent and the reviewing agent traded comments more than 50 times overnight before the author intervened.
The money has moved too. Xu says Anthropic’s internal Claude usage, valued at API rates, grew over 250 times between January and September, and in April it exceeded that month’s personnel costs without counting stock compensation. By mid-August, OpenAI researchers were using a median of more than $600 a day in inference at API prices, and the 90th percentile topped $7,000. She also cites Jellyfish data on 7,548 engineers: the top fifth by token use merged about twice as many pull requests as the bottom fifth, for roughly ten times the spend.
Her one example outside the labs is Spawn, a multiplayer game company. According to founder Jacob Sansbury, the roughly ten-person team spent about $3.1 million on tokens last month, around 30 times its payroll. Xu notes that the figures come from the founder and have not been independently checked.
Two caveats frame all of this. Xu is an investor writing about how to value these companies, and she says organization should now count for more when backing one. Her sourcing is also largely private conversations plus public posts, not audited data. She is candid on the limit of her view: about three months out, and past that she does not know.
For anyone budgeting for agents, the sharpest number in the essay is the spread. If two colleagues with identical titles can draw more than ten times as much compute, a headcount plan stops telling you what a team costs.
Reported by Hannah Xu on X on 30 October 2026.
