OpenAI's Jev clone could help the frontier lab stop its swarming agents

It is one party's illustration of a hackathon demo that checks each agent action against its task, blocking high confidence bad ones and flagging the rest. A scenario estimate inside news analysis, not a production benchmark.

Same shape, unknown internals. Altman described a predefined option set for Luna that keeps image understanding and safety, but the piece notes a limited preview with no third-party testing and undisclosed pricing and schema.

That is the framing, not a finding. The Hugging Face incident is cited as something such monitoring might have caught, yet nothing measured shows a Jev based review would have flagged it in time.

TypeSafe's CEO, who points at synthetic data while conceding fast and cheap is easy, with an intelligence per dollar goal instead. That is a founder assertion in an interview, and the article offers no published eval behind it.

Sam Altman revealed it in an aside at Dev Day. It gives the lab's Luna model a predefined set of options to choose between, such as image categories or different agent behaviors, and he said focusing the model on that choice makes it extremely fast while keeping image understanding, broad language support and safety protections.

OpenAI released it as a limited preview, and the article says TechCrunch had not then spotted developers running it through its paces, though there was clearly interest on X. TypeSafe did not respond to TechCrunch's questions about the new product.

Diogo Almeida joked on X about the beginning of the clone wars and said OpenAI's interest could be a sign that building in a System One compatible way is the future. He also said fast and cheap is easy, comparing it to dice, and that his North Star is always pushing the intelligence-per-dollar Pareto curve.