On October 15, Liquid AI launched d1-ultra, its closed-source flagship decision model. It is the second move in the d1 playbook: on October 7 the company open-sourced the edge-sized d1-3B and d1-omni-600M, and eight days later put a commercial closed-source version on the Liquid API, with Vercel and OpenRouter integrations promised soon. The timing is deliberate — after Jev-style decision models went viral in September, OpenAI and Microsoft piled in, and Liquid AI has now pushed the "no long-form text, just verdicts" track up another notch.
65 points on top, $0.042 per million tokens
According to the official Liquid AI blog, d1-ultra scores 65 on the newly established Artificial Analysis Decision Index, leading every other decision model by at least 8 points; figures cited by Chinese outlet ic.work show Jev 1.13 at 58 and Microsoft's comparable model at 54. Liquid AI says d1-ultra beats its predecessor d1 and Jev 1.14 across 112 unique evaluations, and was specifically trained for calibration — the team found most decision models are overconfident, while d1-ultra produces more trustworthy probabilities. Pricing is input-only at $0.042 per million tokens, with zero output tokens. For reference, the open-weight d1-3B already answers in 8 ms on an RTX 4090; the closed-source version lifts decision quality another tier.
Cheap tokens, expensive questions
Zero output tokens does not mean free. The official pricing docs state that images count at 1.5 tokens per 32×32-pixel patch — a 1024×1024 image costs 1,536 tokens — and, crucially, "each question is billed as its own prompt, including its text and all images." ic.work's analysis did the math: hang 3 quality-inspection questions on one image, and that image's 1,536 tokens get billed in full three times. A single request takes at most 8 images and 10,000 patches — an extremely low unit price multiplied by the question count can make bills rebound violently in composite business scenarios.
The open-weight reality and the generalization cliff
The open-source side tells a more sobering story. ic.work notes that d1-3B averages just 74.1 across 11 public image benchmarks, only 0.2 above its LFM2.5-VL-3B backbone; the 587M d1-omni-600M manages only 15.95 on the Decision Index, with mutually exclusive modalities — essentially an early checkpoint rushed out. More alarming is the generalization risk of the "one forward pass, one verdict" design: in engineering tests, a prototype of this architecture scored 77.3% in-domain but crashed to 54.6% on unseen out-of-domain distributions, close to a coin flip; comparable competitors reach a calibration error of 0.098 on Banking77 and still confidently output 0.9-plus on uninformative inputs. An autoregressive model that misspeaks can walk it back; a single-pass model that misses, misses.
So what
The d1 lineup's positioning is actually clear: work as the millisecond triage clerk at the agent's front door — which team gets this ticket, whether this content is risky — one forward pass, one answer, while long-context reasoning stays with the generative main model. That division of labor makes sense. But if the $0.042 price tag tempts you into letting a single-pass decision model take over an industrial control loop, remember: it generates no tokens, and it generates no second chances. Open-source background in the Hugging Face blog.