Jev gains popularity as a faster, cheaper alternative to major AI models

Most AI companies are racing to build bigger, smarter language models. TypeSafe AI decided to build a faster, dumber one. That’s not an insult. It’s the whole point.

Jev, the startup’s new AI model, launched on September 15 and immediately attracted so much developer interest that it overloaded its own API. The model doesn’t write poetry or generate code. Instead, it makes structured decisions, things like classifying emails, routing support tickets, or flagging transactions, at speeds and prices that make traditional large language models look like hiring a surgeon to apply a Band-Aid.

The speed and price gap is hard to ignore

Jev processes requests in 70 to 500 milliseconds, with a median response time hovering around 0.44 to 0.48 seconds. The pricing is even more striking: $0.042 per million input tokens, with outputs provided entirely free.

In benchmarks conducted at Vercel, Jev proved 5 to 18 times faster than certain OpenAI models on specific classification tasks. For email classification, it came in at 10 to 20 times cheaper than Google’s Gemini model.

TypeSafe AI describes it as a “System One model,” borrowing the term from behavioral psychology. System One thinking is the fast, automatic, intuitive kind. System Two is the slow, deliberate, analytical kind. Most large language models aspire to System Two. Jev is built for System One.

That means it excels at executing fast, structured probabilistic decisions over well-defined questions. If your workflow involves routing millions of inputs through a decision tree, Jev is designed to handle that at scale without burning through your compute budget.

The team behind the model

TypeSafe AI was founded in 2024 by Diogo Almeida, Erik Gafni, and Sasha Sheng. Almeida spent roughly four years at OpenAI, where he contributed to some of the lab’s most consequential projects: RLHF (reinforcement learning from human feedback), InstructGPT, ChatGPT, and GPT-4.

The startup secured $40 million in seed funding led by DCVC, the deep-tech venture capital firm.

Jev is trained exclusively on synthetic data using a method TypeSafe calls Reinforcement Learning for Calibrated Decisions, or RLCD. The goal is to produce outputs that are highly deterministic and reliable, not creative or surprising.

Early feedback from developers has highlighted the reliability of Jev’s confidence scores, which is a critical feature for automation pipelines. If a model tells you it’s 95% confident in a classification, you need that number to actually mean something. Unreliable confidence scores are one of the persistent headaches with repurposing general-purpose language models for production automation.

What Jev tells us about where AI is heading

Jev is closed-source, which has already sparked discussions in the developer community about whether open-weight “Jev-type” adapters might emerge to fill similar niches.

Some early tests indicated Jev doesn’t always outperform on raw accuracy, which means production teams will need to evaluate the speed-cost-accuracy tradeoff for their specific use cases.

The API overload following launch suggests developers are already doing exactly that evaluation.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.