The claim
If general models take over more of the bounded work in a research day, the research organization itself speeds up; the open-ended remainder is the last thing left to solve.
What has to hold for this route to work.
The real bottleneck. The bottleneck is framed as the gap between bounded problems and open-ended research, not as any single missing capability. OpenAI supplied this judgment itself: FrontierScience shows high scores on Olympiad-style questions and low scores on open-ended research tasks. The response is not to add a domain model but to hand agents every bounded part of the research day first.
What is established. With a human watching, agents can already carry definable work through a research day, and they now reach real instruments. By mid-August, 3.1 agent-workdays for every human workday; at MIT, Codex connected the model to laboratory software so agents chose parameters, ran qubit measurements, analyzed results and set the next measurement. The first number is self-reported; the second has a named collaborator.
What is still a claim. Frontier mathematical discovery is still a claim, and this time even authorship of it is unsettled. The Navier–Stokes result was announced by OpenAI itself; Buckmaster and Alpöge say earlier unpublished work had already reached OpenAI, and Bubeck disputes parts of that account while apologising for his wording. The productivity figures are likewise unaudited. A capability claim now arrives bundled with a provenance question.
The chosen product abstraction. The product abstraction is the intern, not the instrument: one general agent plugged into the surfaces scientists already use. Codex reaches into laboratory software; Prism reaches into LaTeX, citations and drafting, free to ChatGPT users. What is sold is a horizontal subscription and API, not a database, model weights or lab services, at the cost that every scientific claim is vouched for by the same channel that makes it.
