Liquid AI's d1-ultra beats Jev 1.14 across all eight categories

Liquid AI built d1-ultra on a larger open-weights base than its earlier d1 model, aiming to handle complex situations where a decision requires reading large amounts of information. The company says the biggest gains came on knowledge-heavy tasks, which is the kind of work where simple classifiers often fall short.

Beyond raw accuracy, Liquid AI stresses the reliability of the probabilities the model outputs. Ideally, when the model says an answer is 90% likely, it should be right about 90% of the time; older decision models tended to output probabilities higher than their actual accuracy. d1-ultra's measured gap between stated probability and real accuracy was 1.5 points, narrower than Jev 1.14's 2.7 points, suggesting its confidence signals are more trustworthy.

The model is available now through Liquid AI's API, with input priced at $0.042 per 1M tokens and no charge for output. The company says it will keep adding features, improving decision accuracy, and shortening response times across decision models of various sizes.