From documents to audio, Liquid AI expands the scope of its decision models
The d1.2 family combines extraction, classification, and confidence estimates. Its Ultra version tops Artificial Analysis’ Decision Index.
Automating document processing means retrieving usable information, identifying where it appears, and deciding whether it needs verification. These are the operations Liquid AI is targeting with d1.2 and d1.2-ultra. The two new decision models accept text, images, and audio, returning structured answers for classification, extraction, and information localization.
In its announcement, the company adds five task types while maintaining input-only billing. The Ultra version also takes first place among decision models on the Artificial Analysis Decision Index. This combination targets repetitive workflows where the answer’s format and the reliability of its probability matter as much as its content. Answers linked directly to the source data
Extraction can target text, a number, or a date. For an invoice, for example, the API returns the requested amount alongside the passage it came from, with a field indicating whether the information was found. Another function retrieves every passage of a specified kind in a document, such as personal data, while preserving its position.
For images, answers can take the form of bounding boxes around an object or multiple instances, accompanied by a count. A numerical estimate can also include an interval. Classification now supports multiple answers at once: a support message can concern both billing and delivery, with a probability assigned to each category.
Audio can be either the material being analyzed or the question itself. The examples include identifying a language or detecting a siren. In its internal tests, the company reports 95% to 97% accuracy on recordings from ESC-50 and FLEURS, with the latter covering 43 languages. The questions are closed-ended, offering three to six choices or a yes-or-no answer: these results do not measure general understanding of unrestricted conversation. An external ranking, separate internal tests
With a score of 75, d1.2-ultra leads the Decision Index at the time of publication. The d1.2 version scores 69. The index combines six evaluations covering documents, decisions, tool calling, and knowledge questions. Its scoring also accounts for probabilities: a confidently wrong answer receives a greater penalty.
The new tasks are measured separately. Without a public benchmark covering them, the company uses held-out test splits from public datasets. For reading values from receipts, it reports 98% accuracy for Ultra and 96% for d1.2. When selecting every object present in an image, the order reverses: 87% for Ultra and 90% for d1.2. The more expensive version therefore does not lead on every evaluated task. Input-only pricing
Both versions are available through the Liquid API, at $0.02 per million input tokens for d1.2 and $0.042 for d1.2-ultra. No output tokens are billed. According to the documentation, a request containing multiple questions is charged for its shared content only once; each question adds its own tokens. Availability through OpenRouter and Vercel is planned, as is a future release of these two versions’ weights, without a specific timeline. Confidence needs to be verifiable
To route a decision for human review, the returned probability needs to match the actual frequency of correct answers. An 80% confidence level is well calibrated if answers assigned that probability are correct roughly eight times out of ten. Calibration describes the reliability of probabilities, rather than the overall accuracy rate alone.
Across 13 benchmarks and 22,505 questions answered by all five models, Liquid AI measures a calibration error of 0.8 percentage points for d1.2-ultra and 1.2 percentage points for d1.2. These figures come from its own evaluation protocol. Within a processing workflow, the probabilities can be used to set a threshold for automating some cases and sending others for human review.