Liquid AI releases d1-ultra and takes first place on the Artificial Analysis Decision Index
Liquid AI released d1-ultra on 15 October 2026, two weeks after its first decision model, d1. It is a hosted model on the same /decisions/v1/systemone endpoint: send text, images or both, ask noul, choice or score questions, and get probabilities back with no generated tokens. It costs $0.042 per million input tokens and was on Vercel AI Gateway and OpenRouter the same day. Liquid says it is built on “a larger open-weight model” with its own customizations; the weights and base model are not public.
The independent number. On 16 October Artificial Analysis put d1-ultra first on its Decision Index v1.0 with 65, ahead of Jev 1.14 (57) and GLiDE and Microsoft-Decision-1 (54). AA explains where the lead comes from: d1-ultra says “I don’t know” on 85% of the knowledge questions it can’t answer, against 2% for d1, while its raw accuracy on that test fell from 41% to 32%. AA also measured 92 ms per task (third fastest, behind Celeris-1 Decision and Jev 1.14) and $32 per million tasks.
Liquid’s own claims (UnverifiedClaim). Liquid says d1-ultra leads all eight categories of a 112-evaluation suite against d1 and Jev 1.14, with the lowest calibration error of any model it tested. The suite and the calibration method aren’t published.
On this site. d1-ultra gets its own page at 5.4 maturity / 7.9 capability / 25 adoption. Capability is higher than d1’s (6.5) because the AA board is independent evidence and AA measured its latency. It still trails Jev on calibration evidence and has a much shorter context (32K).