A new AI model that answers in numbers rather than sentences is drawing attention from AI leaders because it says how sure it is.
San Francisco start-up TypeSafe AI released Jev in early access on 15 September.
It returns a verdict and a confidence score instead of paragraphs of text.
Asked to sort a customer support ticket, it assigns a percentage confidence to each category, such as a billing problem or suspected fraud.
That targets a known weakness of large language models (LLMs), the systems behind ChatGPT and Claude: they can present wrong answers with total confidence.
Software can act automatically on answers above a set confidence level and pass the rest to a person.
TypeSafe was founded by Diogo Almeida, who spent four years improving ChatGPT's responses at OpenAI before leaving in 2024.
The company emerged from two years of secret development with $40 million of seed funding led by DCVC, a venture capital firm.
It charges $0.042 per million tokens of input, the units of text the model reads.
Output is free, and TypeSafe says Jev answers in under half a second.
That suits it to high-volume work such as classifying millions of emails.
Some have compared Jev with DeepSeek, the Chinese lab whose low-cost models unsettled Silicon Valley.
Developers mostly see it as a complement to LLMs rather than a replacement.
In one likely set-up, Jev sorts a customer's message into a pricing or refund query, then passes it to an LLM to write a human-sounding reply.
