Arcturus Labs Argues OpenAI Could Replicate Jev's Core Trick

Arcturus Labs founder John Berryman argued Monday that OpenAI is positioned to replicate the classification technique behind TypeSafe AI's Jev model, a week after Jev became the fastest-adopted model in Vercel's AI Gateway history.

Berryman's post, published Sept. 21 on Arcturus Labs' blog, contends OpenAI has for years used its language models as implicit classifiers, and that formally training and packaging that capability as a standalone product would let OpenAI reproduce, then extend, what Jev already does.

What Berryman actually claims

Berryman pointed to his own earlier writeup showing GPT models already make binary tool-calling decisions internally, through a special token, to=function., embedded in the model's ChatML markup. He argued that if OpenAI trains a model the way TypeSafe trained Jev, extracting calibrated probabilities from a single forward pass instead of generating text, OpenAI could fold that ability into its existing models for task routing, reasoning-trace verification and safety screening of tool calls before they run.

The adoption number driving the argument

Jev reached 13 percent of paid teams on Vercel's AI Gateway within 24 hours of its Sept. 15 launch, twice the first-day share of the GPT-5.6 family and six times Claude Fable 5.1's, according to Vercel's Sept. 18 announcement. Vercel said Jev passed every comparison model within its first 12 hours, though the company cautioned that "its first-day adoption was unmatched among recent launches; the next test is whether that early adoption lasts."

Comparison	Jev's 24-hour adoption multiple
Jev vs. GPT-5.6 family	2x
Jev vs. Claude Fable 5.1	6x
Jev's share of paid AI Gateway teams	13%

Source: Vercel AI Gateway launch data, accessed 2026-09-22.

TypeSafe's own answer on Jev's moat

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, says the durable part of Jev is not its architecture but the process it uses to turn training data, all of it synthetic, into calibrated probabilities, a distinction the prior-art dispute over Jev's origins already put on the record this month. Almeida, TypeSafe's chief executive, told Latent Space in a Sept. 21 interview, "if model quality matters, then we are gonna be in a very good position for a long time," describing TypeSafe as "a data lab rather than a model lab."

What would prove Berryman wrong

Berryman's own post lists the case against his argument. Jev works best for what he calls System One judgments, snap classifications, not math or multi-hop reasoning, and TypeSafe's own documentation describes what it calls "Jev's jagged edges" in domains the company has not smoothed out. Berryman wrote that he has found domains where Jev's probabilities do not hold up, without naming which ones. Neither OpenAI nor TypeSafe has said publicly whether OpenAI is building the classifier product Berryman describes, and the test he proposes, whether Jev's adoption curve keeps climbing once developers move past a first trial, will not resolve either way for several more weeks.

The argument lands as OpenAI is already shipping GPT-6 Astra into narrower product lines faster than outside evaluation can keep up, a pattern a robot-arm safety benchmark also flagged this month when the model completed 60 of 100 instructions researchers had marked as dangerous. Berryman's post does not say when, or whether, OpenAI has started building the classifier he describes.