Ex-OpenAI Researcher Launches Jev, an AI Model That Scores Options Instead of Generating Text

A model that scores instead of writes

Startup TypeSafe AI has introduced Jev, a model designed to deliver narrow judgments and probabilities inside software rather than generate chat responses, emails, or code. Co-founder and CEO Diogo Almeida previously worked at OpenAI and is listed among the authors of the InstructGPT research that underpinned ChatGPT's development.

The core mechanism: developers define a question and a fixed set of possible answers, and Jev returns a score or probability for each option. It does not produce free-form text.

How it works in practice

TypeSafe's documentation illustrates the intended use with a customer support example. A shopper writes to an online store claiming they were charged twice and requesting a refund. Before a human or another system handles the message, it needs to be classified — is this a payment issue, a delivery issue, or a return? Does the customer want money back or just an explanation?

Developers connect their software to Jev and define these categories in advance. The model returns a label — for example, "payment issue" — along with a probability that the customer wants a refund. The calling application then applies its own rules: payment issues route to accounting, refund requests get flagged, and unclear cases go to a human reviewer. Jev does not verify whether a double charge actually occurred; that determination still requires checking real transaction data, separately from the model's output.

Speed claims and pricing

According to TypeSafe, Jev responds in 70 to 500 milliseconds, which the company says is many times faster than current large language models. TypeSafe attributes this to skipping step-by-step text generation and computing multiple outputs in parallel, meaning additional questions in the same call add little latency, according to the company.

TypeSafe lists a price of $0.042 per 1 million input tokens and says output tokens are free, since the model does not generate extended text.

TypeSafe suggests additional uses beyond support ticket sorting: detecting buying intent, tagging incoming requests by topic, deciding when a human should take over a conversation, or checking a chatbot's draft reply against conversation history before it is sent — for example, verifying it doesn't reference a refund that isn't recorded in the customer's account.

Limited independent verification

TypeSafe's published benchmarks compare four workflows the company built itself, using outputs from other AI models as reference points rather than independently verified correct answers. GPT-6 Astra is not included in the comparisons. Standard large language models already support constrained, structured output — OpenAI's Structured Outputs feature covers similar ground — so a fixed-format response alone does not differentiate Jev from existing tools. Its advantage would need to come from speed, cost, or judgment quality at scale.

TypeSafe markets Jev as unable to hallucinate, but this claim only applies to output format: the model cannot produce an answer outside the predefined set of options. It can still select the wrong option among those allowed. Whether Jev's judgments are accurate enough for a given task is left to individual companies to test.

Access to Jev is currently limited to a developer waitlist.

What this means

Jev targets a real gap: many production AI pipelines don't need fluent prose, they need fast, cheap classification decisions buried inside larger systems — routing, intent detection, pre-send validation. Skipping autoregressive generation for a scoring-only architecture is a legitimate way to cut latency and cost for that narrow job. But the company's own benchmarks rely on self-built workflows and model-generated reference answers rather than ground-truth validation, so quality claims remain unverified. The bigger question is whether structured-output modes on general-purpose models, which already support fixed categories, can match Jev's speed and price without requiring a separate model and API integration. Until independent benchmarks appear, Jev's value proposition rests largely on TypeSafe's own numbers.