Liquid AI released two updated decision models on November 16, and the larger one, d1.2-ultra, now sits first among decision models on the Artificial Analysis Decision Index v2.1. According to the company’s blog post, it scores 75, two points above the previous best. Both versions are live on the Liquid API.

A decision model is a small, specialized system that does not write prose. You hand it an input and a question, and it returns a structured answer along with the probability it assigns to that answer. Think of it as a classifier or extractor you call over an API, with a confidence number attached to every result.

Five new tasks widen what that answer can look like. The first reads a value (text, number, or date) from a document, and returns “not found” if an invoice never states it. The second finds every passage of a given kind, such as each name, phone number, and email address in a note, with positions. The third locates objects in an image, returning one box or a box per instance plus a count. The fourth estimates a quantity with a range, for example a best guess of 12 people in a photo with a spread of 10 to 14. The fifth picks all options that apply, such as billing and delivery for a support message, each carrying its own probability.

Audio is the other addition. A recording can be the subject of the question, such as which language a speaker uses or whether a siren sounds behind them. The question can also arrive as speech instead of text. Liquid reports that both models score 95% to 97% on recordings kept out of training when asked to name a sound (ESC-50) or a spoken language (FLEURS, spanning 43 languages).

The ranking rests on six benchmarks. Those span extraction from resumes, receipts, and research papers, plus tool calling and difficult knowledge questions. The index also grades the probability each model attaches to its answers, so confident mistakes cost more than hesitant ones. Liquid adds that nothing beats d1.2 at a lower price, and no rival tops d1.2-ultra whatever it costs.

Calibration is where the release makes its sharpest case. When a model says 80%, it should be right about four times in five, which lets a team send low-confidence items to human review. Liquid compared five models over 13 benchmarks and 22,505 shared questions. The d1.2-ultra gap between predicted and actual accuracy is 0.8 points, against 1.2 for d1.2, 2.7 for Jev 1.14, 3.7 for Perplexity’s Decider v1.3, and 7.2 for GPT-6 Luna on its Decisions API.

The caveats sit in the post itself. Liquid ran the five-model calibration comparison, and no public benchmark covers the new tasks yet. For those, it built internal test sets from reserved portions of public datasets. Scores there run from 98% for d1.2-ultra reading receipts (CORD, SROIE) down to 87% for it listing every object in a photo (COCO), where d1.2 actually does better at 90%. On personal data, d1.2-ultra reaches 94% precision and 93% recall against 92% and 88% for d1.2.

Input tokens are the only billed item: $0.02 per million for d1.2, $0.042 per million for d1.2-ultra. Images are counted at 1.5 tokens for each 32x32 patch, which makes a 1024x1024 picture 1,536 tokens. Extra questions about the same picture add only their own text. Liquid says OpenRouter and Vercel availability is coming soon.

For teams running document intake or moderation, the number to test is calibration, not the headline score. Pull a sample of your own invoices or support tickets, check whether the 80% answers are right four times in five, and set your human-review threshold from that.

Liquid AI, “d1.2: State-of-the-Art Omnimodal Decision Models,” Liquid AI blog, November 16, 2026.