A specialized decision model from Liquid achieved 100% accuracy on a financial verification benchmark, equaling the performance of general purpose models including GPT-6.1 Sol and the Claude 5.5 family. The d1-ultra model correctly identified all 1,960 claims on the ClaimBench test, which requires AI to determine if 10-K corporate excerpts support, contradict, or ignore specific statements. It costs $0.21 per 1,000 claims and responds in under a second, making it roughly 12x cheaper and 5x faster than Haiku 5.5, the most affordable general model to hit the same score.
Vals AI launched the ClaimBench benchmark on October 20, initially testing 15 models to evaluate trust in decision-specific AI. While Perplexity’s Decider v1.2 led initial subsets with a score of 67.1%, d1-ultra is now the only decision-oriented system to match frontier LLMs. This version improves upon Liquid’s previous d1 model, which missed 26 claims due to subtle contradictions in wording, although the newer d1-ultra costs about twice as much to operate as the original.
