"193 Times Faster and 444 Times Cheaper"… Why Developers Are Raving About AI 'Zeb' That Can't Write a Sentence

JAVE achieves the fastest paid user adoption in history

No need for sentence generation, code writing, or reasoning explanations

Developers excited by rapid decision-making and low cost

Unlike large language models (LLMs) like ChatGPT, which generate answers by predicting words one at a time, a new artificial intelligence (AI) model has emerged that doesn't produce sentences, but instead presents only judgments and probabilities. This model is receiving keen interest from developers as it is specialized for handling repetitive decision tasks, and it helps reduce unnecessary token costs and hallucinations that can occur during text generation.

A comparison of how Zeb and existing large language models operate. Screenshot from the TypeSafe AI website.

According to industry sources on September 22, U.S. AI startup TypeSafe AI has released its AI model 'Zeb' to the public after two years of confidential development, making it available for all developers to use.

Zeb was developed by Diogo Almeida, a former OpenAI researcher. Almeida participated in projects such as InstructGPT, which formed the foundation for ChatGPT, and has spent two years developing this technology in secret after establishing TypeSafe AI.

Zeb became the fastest-adopted new AI model in history, with about 13 percent of paid users using it within 24 hours of its registration on the developer platform Vercel AI Gateway. According to Vercel, the proportion of paid users who used Zeb during this period was more than twice as high as when OpenAI's GPT-5.6 was launched. Right after launch, a surge in users even caused service disruptions for a brief period.

The biggest difference between Zeb and existing LLMs lies in the way it produces answers. LLMs generate sentences by predicting the next word based on the user's input, building the response word by word. In contrast, Zeb selects among multiple choices or presents whether statements are true or false—or provides scores—accompanied by probabilities.

Since it does not generate sentences, Zeb incurs lower costs for answer generation. The standard price is $0.042 per one million input tokens, with no additional cost for outputs. Its response time is approximately 70 to 500 milliseconds. According to TypeSafe AI’s own evaluation, Zeb is up to 193.6 times faster and 444.6 times cheaper than conventional LLMs.

These advantages make Zeb well-suited for handling simple, repetitive decision-making tasks that occur throughout software development and operations. By providing only the required judgment values for classification or condition-checking tasks, Zeb keeps token costs low. In environments where AI agents divide work into multiple steps, developers can use Zeb for classification and selection, reserving more complex LLMs for steps requiring sophisticated reasoning, thereby reducing costs.

There are already reports of real-world cases from developers. For example, developer Justin Schroeder revealed that he created an autonomous driving simulator in just one hour by providing contextual inputs to Zeb and having it choose driving directions. This tests to what extent a non-generative model can perform complex tasks through repeated choices.

Zeb’s arrival is also raising the possibility of further segmentation in how AI models are used, moving beyond a generative AI-centric approach. An industry insider commented, “Rather than a general-purpose LLM, it’s closer to a high-speed classifier or judgment function. Since its operating principle is different from traditional AI, its use cases will also diverge.”

This content was produced with the assistance of AI translation services.