Jev, the AI Model That Can't Chat but Is Far Faster and Cheaper Than GPT

Startup TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has officially launched Jev—an artificial intelligence model that is not designed to chat, write emails, or generate code. Instead, Jev only outputs structured decisions that can be executed directly by software or automated agents. The pricing is highly aggressive: $0.042 per million input tokens, with no additional cost for output. For developers who have long paid dearly for large language models just to get a few classification words, Jev offers a tempting proposition.

The model classifies information, assigns scores, and returns probabilities without needing to produce explanatory paragraphs like GPT or Claude typically do. This approach immediately caught the developer community's attention. Pranit Sharma, an engineer at Vercel, reported that replacing a conventional language model with Jev for command security classification produced responses 5 to 18 times faster, with accuracy that actually improved. Other developers found that Jev classified business emails at a far lower cost than Google's Gemini, though Gemini was slightly more accurate.

Initial enthusiasm briefly overwhelmed TypeSafe's API, according to a TechCrunch report. For companies running automated systems with thousands of decisions per hour, this experiment shows how to cut the cost of routine AI tasks without having to rebuild software. However, relying on a third-party model for decision-making still carries operational risk.

The big question is not just about price. The technology industry has spent years perfecting language models that can reason, write, and talk with humans. TypeSafe has instead taken a narrow path: building a model focused solely on fast, structured judgments inside existing software. Jev supports three question types—Choice (selecting from specified options), Score (evaluating on a numerical scale), and Noul (estimating the probability that a statement is true).

TypeSafe calls Jev a "System One Model," borrowing psychologist Daniel Kahneman's terminology about fast, intuitive thinking versus slow, deliberative reasoning. The company claims Jev uses a new architecture and a training method called Reinforcement Learning for Calibrated Decisions (RLCD) designed to make its probability estimates better reflect uncertainty. "Think of Jev as a frontier intelligence function call: unstructured state in, typed probabilistic decision out," Almeida said in the launch announcement.

Despite being fast and cheap, Jev's accuracy has drawn scrutiny. Results published by TypeSafe itself show the trade-off: independent testing by Explainx.ai recorded an aggregate score of 67.8% for Jev, while the strongest comparison model in the company's evaluation reached 74.1%. This accuracy gap deserves attention. Applications that sort low-priority customer messages might tolerate the occasional misclassification. But systems that authorize financial transactions or potentially destructive software commands have very little room for error.

"We have lightning in a bottle, but it's useless," Almeida told TechCrunch, describing his frustration with the limitations of conversational AI for software automation.

Another concern arises from TypeSafe's claim that Jev cannot hallucinate. This claim refers to Jev's inability to produce responses outside the predefined structure—if an application asks the model to choose from five categories, Jev cannot invent a sixth. However, Jev can still pick the wrong category or assign a high probability to an incorrect answer. TypeSafe's guarantee concerns the structure of responses, not the factual accuracy of decisions.

More critically, a VentureBeat investigation on September 21 uncovered an experiment with a potentially destructive computer command. Jev initially gave a 0.76 probability of blocking the command. After an engineer added a fake authorization message to the information presented to the model, its blocking probability dropped to 0.48. This experiment involved only a single command and cannot determine how often such attacks succeed. Still, it shows that harmful information inserted into an application's input can influence Jev's judgment. A security system that approves unauthorized commands does not become more reliable just because its decision is made in 70 milliseconds.

For Indonesia, the Jev phenomenon reflects an important shift in the global AI economy that will sooner or later spill over into the domestic market. Until now, technology companies in the country—from fintech startups to e-commerce—have relied on large language models for various tasks, from automated customer service to sentiment analysis. High inference costs have often been a barrier to large-scale AI adoption. If specialist models like Jev can truly cut costs by hundreds of times, a wave of decision-based automation could reach small and medium-sized businesses that were previously priced out. However, relying on third-party APIs for critical decisions—such as transaction verification or risk classification—demands extra caution, especially regarding data sovereignty and consumer protection regulations.

TypeSafe named Jev after 19th-century economist William Stanley Jevons, whose work examined how greater resource efficiency can actually drive higher consumption. The company hopes a similar effect will occur in AI: lower inference costs will make intelligent decisions viable in software processes that have long relied on rigid rules or human intervention. LangChain, for example, has tested Jev to evaluate the behavior of other AI agents. In a September 20 experiment with five weather-agent responses, Jev matched human binary judgments on 500 repeated decisions, averaging 0.44 seconds and $0.00035 per evaluation. Even so, this test covered only one agent and a limited set of tasks—consistency does not always mean correctness.

Going forward, the most decisive metric for companies considering Jev may not be speed or price, but the cost of wrong decisions. A routing error can send a customer to the wrong department. A security misclassification can authorize a destructive command. These are very different risks, and the same model may not suit both. TypeSafe has indeed made these experiments remarkably cheap. Whether Jev becomes a fixture in enterprise software will depend on what developers find once promotional pricing and early benchmarks give way to sustained production workloads. Can specialist models like Jev replace reliance on general-purpose language models, or will they instead become a complement in an increasingly layered AI ecosystem?