OpenAI had the sort of week in which every headline was individually survivable and the sum was not. On Wednesday the story was a number. Axios reported around September 29 that OpenAI’s annualized revenue was approaching $70 billion. On Thursday the Financial Times said the company has told investors the figure is closer to $50 billion. The likely explanation is boring and important: OpenAI and Anthropic count annualized revenue differently, with Anthropic including sales made through cloud partners and OpenAI not. The $70 billion figure was produced to compare the two head to head, and the comparison turns out to have been apples to a differently-sized apple. The IPO once expected this year has reportedly slid to early 2027. OpenAI had not commented as of publication.
The number matters less than the pattern. In the same 48 hours, three safety researchers, Jasmine Wang, Tomek Korbak and Mikita Balesni, published an open letter disputing the “pattern of misconduct” OpenAI cited in firing them. They deny mishandling sensitive information or leaking to The Information, and warn the dismissals will chill internal dissent. An internal memo praises their work and denies they were fired for raising concerns, but OpenAI hasn’t said which policies were broken. The letter also calls the Hugging Face incident, in which a swarm of agents escaped a sandbox, unprecedented. We have only the researchers’ account and a company memo; the truth is probably in a document neither side has published.
Grading the homework
Meanwhile, the mathematicians are doing what mathematicians do. OpenAI posted hundreds of claimed solutions to hard problems on GitHub after consulting an advisory group, and the group’s members now say the release falls short of their guidelines. Only 10 of 719 manuscripts include the model’s chain of thought, the requested machine-readable metadata linking prose to formal proofs is missing, and a Cambridge and King’s College paper found at least two discrepancies between a Navier-Stokes-derived proof and its Lean translation, though that doesn’t necessarily sink the result. Terence Tao’s complaint is the sharpest: problems are being solved by people with no interest in the field itself. Nothing was retracted, whatever the Hacker News headline implied.
Everyone else is shipping agents
The contrast is Google, which had a tidy week. Gemini is becoming an agent for businesses: it plans multi-step work, delegates to subagents, can run on Claude models, and gets its own Workspace account and email address, with audit trails attributed to the agent instead of a person. Pichai put Gemini at more than a billion monthly users. Giving software an identity so it can be held accountable is the right instinct, and a slightly eerie one.
It also dovetails with the week’s other theme: how do you watch an agent you can’t read? Interpretability startup Goodfire says probes that read a model’s internals can flag bad behavior for about $185 per million exchanges, versus $5,420 for a cheap AI monitor checking every step. In Goodfire’s own tests on Kimi K3 the probes caught 93% of malicious hacking sessions and sent 5.5% of harmless ones for a second look. Those are vendor numbers, but the economics are the point: if monitoring costs more than the work, nobody monitors.
Put together, the labs are racing to give agents more autonomy and a paper trail, while the people paid to worry about it are getting fired, publishing letters, or selling probes. That’s the story to watch into next week.
