What Jev’s fast followers reveal about the AI race

Erik Kokalj, who works on developer experience at US-based computer vision company Roboflow, built a demo that flags eggs on a video feed as clean, dirty, or cracked in one afternoon. He still needs an object detector to find the eggs.

But instead of training a separate model to determine the eggs’ condition, he used OpenAI’s Decisions API, which was rolled out on October 22.

Decisions API belongs to a new category of AI products known as decision models. Unlike traditional classifiers that generally require training for particular jobs, these models can choose from predefined answers without additional task-specific training.

US startup TypeSafe helped popularize the category when it launched Jev in mid-September, sparking a wave of industry interest.

Jev is currently the most popular decision model on OpenRouter, a platform that lets developers call many AI models through one interface. It has processed over 5.8 billion requests over the past month.

Within weeks of Jev’s launch, major AI players including OpenAI, Microsoft, and Cloudflare introduced competing decision models or APIs.

Cloudflare’s Clef, for instance, was trained and launched in just a week, while OpenAI’s Decisions API builds on the company’s existing model weights. Startups in India, such as Sarvam, also released their own decision models.

Jev’s early success shows there’s a huge demand for fast, inexpensive decision models, as developers told us previously. But the speed at which rivals emerged raises a question: If the core tech can be replicated so quickly, what will set these products apart?

Why rivals could move so fast

Jev took two years to build, TypeSafe co-founder and CEO Diogo Almeida said on a podcast. He recalled building an early version in the first week that, in his view, resembled some of the approaches rivals use today. However, he found it unreliable.

Tech in Asia asked TypeSafe for comments but got no response.

Jev has revived a prediction Cloudflare made in 2023, says Michelle Chen, who leads product for the company’s AI platform: AI models would become smaller and faster, making it increasingly important to run them on servers closer to users to cut response times.

The next battleground for decision models

The catch with decision models