Jev, an AI model built for quick decisions, went viral among AI developers and Silicon Valley cognoscenti when the startup TypeSafe AI released it in September. These early adopters were quick to experiment with the AI model, which could perform all kinds of classification decisions—like sorting online customers into buckets based on certain data points or categorizing documents—at a fraction of the cost of many leading AI models. Programmers showed it sorting email, spotting writing mistakes and even playing games. Within a day of arriving on Vercel’s AI Gateway, a service for accessing different models, nearly 13% of its paid teams had tried Jev. Vercel said Jev reached more than twice as many paid teams in its first 24 hours as any previous model launch.
Perhaps the clearest sign that TypeSafe was on to something: on Oct. 6, just three weeks after Jev’s launch, OpenAI rolled out a competing product, called Decisions API.
These services are competing to make AI practical for the thousands of routine judgments businesses might want to automate each day. At that volume, small differences in cost, speed and error rates can change whether automating a task is worthwhile. TypeSafe trained Jev specifically for this work. OpenAI’s Decisions API uses GPT-6 Luna, an existing model in its lineup.
An AI model that skips the prose
TypeSafe CEO Diogo Almeida told Fortune he began thinking about Jev after helping develop the technology behind ChatGPT at OpenAI. Almeida saw a gap between models’ ability to answer people’s questions and their usefulness in automating routine work. Almeida said he wanted to address AI’s “massive over-promise under-deliver issue” and avert an “AI winter.” “Building a company just happened to be the most effective way to do that,” he said.
Before joining OpenAI in 2020, Almeida took a year and a half off to travel, meditate and play games. He says he reached “top-2 in Hearthstone,” the digital collectible card game, an accomplishment he joked might have been harder than building Jev.
Almeida has said he had tried to work backward from a future in which AI had transformed the economy. In that future, would most requests to models come from people, or from code running automatically? He concluded that the vast majority would come from code.
A customer-service program, for example, could ask a model whether an incoming message concerns a billing issue or a technical problem, then use the answer to route it to the appropriate team.
Developers could set a threshold for the model’s confidence about its routing decision. Above that threshold, the message would be routed automatically to billing or technical support. Below it, and it would be flagged for a human to check..
This sort of classification task is something that rudimentary AI, or machine learning systems, have handled for decades. Computer science students are often taught to build these classifiers in machine learning courses. Email services have long used these simple models to filter spam. A model can estimate how likely a message is to be spam, and the software can move it to the junk folder when that estimate crosses a threshold set by its developers.
The problem with these simple models is that they tend to be relatively inflexible, without much understanding of a document’s text, and they require some technical know-how to create.
In some ways, large language models (LLMs), which are what today’s best known AI systems are based on, represented an improvement on these simple classification systems. LLMs can perform classification tasks too, sorting customer-service messages based on written descriptions of categories such as billing questions or technical problems. They have a much deeper understanding of the text they are processing, so can make finer-tuned decisions. Developers can also change the LLM’s instructions to use the same model for other jobs. So that makes these systems more flexible than the old classification models.
But LLMs take more computing power to process these decisions, which makes them both more expensive and slower than the old, simple machine learning classifiers. When the LLM has to handle thousands of these requests throughout the day, the cost and delay of each call add up.
TypeSafe says Jev is designed to handle these judgments faster and more cheaply than general-purpose language models, but with the same ability to set up the classifier through easy, natural language instructions that LLMs offer. Developers describe what they want Jev to evaluate, and Jev returns choices, scores or probabilities their software can use. The company says it produces those outputs together, avoiding the time spent generating an answer piece by piece. Developers still have to establish whether its judgments are accurate enough for their particular task.
Almeida initially thought making the models behind Jev reliable would take him a week. Instead, TypeSafe spent two years developing Jev. “Making the models reliable was so much harder than expected,” he told Fortune.
What developers love about Jev
Niels Mouthaan, the independent developer behind the Daily time-tracking app, told Fortune he tested Jev in a spelling and grammar prototype that gave users feedback as they typed. Users expect “close-to-instant feedback while typing,” he said. He was impressed by Jev’s speed and saw promise in using it to flag errors, with a language model suggesting corrections only where needed.
TypeSafe calls Jev a “System One” model, borrowing the term for fast judgments from psychologist Daniel Kahneman. After experimenting with Jev, Ashutosh Mathore, a founder of environmental compliance software startup Atlensa, told Fortune its speed made him reconsider “where we may be using LLMs out of habit even when the task is not really a language problem.”
Jev is named for economist William Stanley Jevons, Almeida wrote in TypeSafe’s launch post. Jevons argued that more efficient steam engines would, somewhat counterintuitively, increase overall coal consumption by making coal-powered work cheaper. The lower cost of coal-power would increase demand, leading people to use it for more things. Applied to AI, the idea is that lower costs could make many more tasks worth automating.
Paolo Rosson, head of applied AI at bookkeeping software company Dext, said that the problem with LLMs is that “you pay for a full written answer when all you need is a label.” In a personal project screening proposed software changes, he has used Jev to reduce how often routine changes go through a larger model. He estimated the cost of Jev’s model calls at seven cents per 1,000 changes. He also spent considerable time deciding which probability estimates should trigger human review, depending on “how much risk you’re ok with,” he told Fortune.
Still, for developers now accustomed to using LLMs, switching to Jev takes some getting used to. “To get successful outcomes from Jev you need to shape questions and prompts different from traditional LLMs.” Aaron Roy, the developer of ingredient-scanning app GlutenOrNot, said. In an email-sorting experiment, Roy asked Jev to flag emails that needed his attention. It let important messages slip through. He then spelled out what mattered to him, asking whether an email concerned his money, children or medical issues. Those more specific questions reduced the misses in his test.
Then OpenAI showed up
While still at OpenAI, Almeida wrote a document outlining his idea for models optimized to work with software and showed it to OpenAI CEO Sam Altman. Altman urged him to put aside his other work and pursue it, Almeida recalled in a LinkedIn post. “He told me to stop working on RLHF and work on my idea instead,” he wrote. Eventually, Almeida left to found TypeSafe. But OpenAI never rolled out its own classification engine until Jev showed up and caught fire among developers.
OpenAI’s Nikunj Handa credited Jev with “inspiring this whole thing” in an interview with the podcast “Latent Space.” The Decisions API had not been on the development roadmap four weeks earlier, he said. For the first version, the team used GPT-6 Luna without further training, constrained it to return structured answers, processed multiple questions at once and tuned the system to deliver the first decision faster.
The company describes the service as a way for applications to choose “the right model, tool, or action” in near real time. OpenAI says its dedicated interface can return decisions up to ten times faster than calling GPT-6 Luna through its general-purpose Responses API. Like Jev, it can select from preset options, estimate whether a condition is true and assign scores. It also accepts images, allowing an application to check a photograph or browser screenshot. Jev currently accepts only text. OpenAI lists an input price of $0.10 per million tokens, the small units of text models processed, compared with Jev’s $0.042. Neither service charges for output tokens.
Besides its deep pockets and large user base, OpenAI may have certain advantages in this contest with Jev. For instance, Roy told Fortune he is running Jev alongside Anthropic’s Claude on real users’ ingredient scans, with Jev evaluating the ingredient text obtained from those scans. He plans to use Jev for the initial classification, turning to “slower, more expensive frontier models whenever Jev is unsure about how to classify the inputs.” In these kinds of use cases, developers who already use OpenAI for more complex tasks might prefer to now obtain those preliminary, fast classification judgments from the same provider.
On the other hand, James Hardiman, a general partner at DCVC, which led TypeSafe’s $40 million seed round, points to the startup’s two years of development, particularly its work on reliability, as an advantage over its larger rival. “There’s this kind of intangible element I think to some of these things that the benchmarks don’t capture, and that’s why I think actually in production, Jev will perform better than these kind of fast follow clones that we started to see come out,” he told Fortune.
TypeSafe says it has developed a new model architecture and a training method intended to make Jev’s probability estimates reflect how often its judgments are correct. While it’s uncertain whether Jev can go the distance against OpenAI and other competitors who are also likely to jump into offering similar decision AI products, it’s already clear that Almeida’s intuition about the need for a better way to set up classification systems was correct. He spotted the wave early. Now OpenAI and everyone else is paddling to catch up.
The post Jev, an AI for making quick decisions, has been a viral hit in Silicon Valley. But OpenAI is hot on its heels appeared first on Fortune.
