Liquid AI released d1.2 and d1.2-ultra, updated versions of its decision models that can take text, images and audio as input. The company says d1.2-ultra scored 75 on the Artificial Analysis Decision Index v2.1, the highest score among decision models and two points above the previous leader.
Both models are available through the Liquid API. Liquid AI lists d1.2 at $0.02 and d1.2-ultra at $0.042 per million input tokens. The company says both bill only for input tokens; its announcement does not specify the unit for d1.2’s listed price.
What can the models do?
The update adds five API tasks: reading a value from a document, finding every passage matching a category, locating objects in images, estimating object counts, and selecting all applicable choices. For example, a model can extract an amount from an invoice and return “not found” if it is absent, or draw a box around a signature. The API can also return probabilities for possible answers.
Audio can be the material being analyzed, such as a siren in the background, or the question itself: a user can ask something aloud rather than typing it. The new answer types use the same API endpoint and request format as the existing tasks. Liquid AI says OpenRouter and Vercel support is coming, but gives no date for that availability.
What do the scores show?
Artificial Analysis’s index tests six areas, including decisions and information extraction from resumes, receipts and research papers, as well as tool use and difficult knowledge questions. It also considers how well a model’s confidence matches its accuracy, so confidently wrong answers count against it.
Liquid AI reports that d1.2-ultra had a calibration error of 0.8 points across 13 benchmarks and 22,505 questions, compared with 1.2 for d1.2 and 7.2 for GPT-6 Luna on its Decisions API. For the new tasks, Liquid AI tested internal sets made from held-out public datasets; it reports 94% accuracy for d1.2-ultra locating described objects in photos and 93% for estimating counts within one. The company also reports 95% to 97% accuracy on audio questions about sounds or spoken languages, tested across 43 languages. These results are company-reported; the announcement says no public benchmark yet covers the five new tasks.
Why it matters
People building document or image workflows can use the models to extract fields, flag missing information or send uncertain answers for review. The reported scores suggest useful capabilities, but the new tasks lack a public benchmark, so performance on a particular user’s documents or images remains unknown.