OpenAI launches rival classification product three weeks after Jev's viral debut

Classification is not new work. Email services have run simple models to flag spam for decades, and these classifiers turn up as a teaching exercise in machine-learning courses. ^[14] The old models are cheap but rigid, with little grasp of a document's text, and they need someone technical to set them up. ^[15] Large language models do the same sorting with a deeper read of the text, and a sentence of plain instruction can repoint them at a new job. ^[16] But each LLM call burns more compute, so at scale they cost more and run slower, and over thousands of requests a day that adds up. ^[17] Jev's pitch is to combine the two: the flexibility of a language model at something closer to the price of the old classifier. ^[3]

When a customer-service system asks a model whether a message is a billing question or a technical one, thousands of times a day, a cent or a few hundred milliseconds per call compounds into whether automating the task pays at all. ^[9][18] TypeSafe says it trained Jev for exactly this. ^[2] OpenAI's answer, shipped October 6, runs on GPT-6 Luna, a model already in its lineup, wrapped in a new Decisions API. ^[7][8]

OpenAI did not train a bespoke model to meet a three-week-old startup; it repackaged one it already had. ^[8] The move validated TypeSafe's category, and OpenAI did it with distribution a startup cannot match.

Neither company has published a price per decision or an error rate, and those are the two figures that decide which product wins the workload. The one measured signal so far is trial: within a day of arriving on Vercel's AI Gateway, nearly 13% of paid teams had tried Jev, and Vercel said Jev's first-day count of paid teams was more than double that of any earlier launch. ^[5][6]

Jev could undercut a general model on cost per decision by enough to keep the high-volume jobs. Or GPT-6 Luna proves good enough and OpenAI wins by already sitting in the stack. A third outcome is that the price of a single classification collapses toward zero. That would hurt a one-product startup more than a company selling an entire model lineup.

The viral launch proves there is demand for cheap classification. Whether TypeSafe captures that demand is a separate question, and the launch does not answer it. OpenAI matching the product in three weeks with a model it already had suggests the barrier to entry here is low. I would change that view if Jev's trained-for-the-task economics beat a general model by a margin wide enough that developers pay to move work off OpenAI's platform.

Almeida, who helped build the technology behind ChatGPT at OpenAI before founding TypeSafe, worked backward from a future in which AI had reshaped the economy and concluded that most requests to models would come from code, not people. ^[10][13] He told Fortune he wanted to fix AI's "massive over-promise under-deliver issue" and avert an "AI winter." ^[11] "Building a company just happened to be the most effective way to do that," he said. ^[12]