d1.2 / d1.2-ultra
Intro
d1.2 and d1.2-ultra are the new Liquid AI decision-model versions formally released on 16 November 2026, adding audio input and five new tasks on top of the existing d1-ultra line and earlier d1 models. @liquidai posted the introduction at 2026-11-16T15:08:30Z (23:08:30 Beijing time), beginning “Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models.” The post says both models can process text, images, and audio, with new output capabilities, and that d1.2-ultra scores 75, first on the @ArtificialAnlys Decision Index.
The official blog at liquid.ai/blog/d1-2 is titled “d1.2: State-of-the-Art Omnimodal Decision Models.” It describes them as new versions of Liquid’s decision models, now supporting audio input and five new tasks: reading a value from the input (text, number or date — for example, the amount due on an invoice, returning the characters it read, or “not found” instead of guessing if the value is absent); finding every passage of a kind in a text (spans, such as every piece of personal data in a note, with its position); locating things in a picture (one box, or a box for every instance plus a count); estimating a number with a range; and picking every option that applies (choice with multiple set to true). Audio can be what you ask about (for example, whether there is a siren in the background, or which language is spoken) or the question itself, spoken instead of typed.
Liquid says d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, first among decision models and two points ahead of the previous best. The index covers decisions and extraction on résumés, receipts and research papers, tool calling, and hard knowledge questions, and also scores the probability a model assigns to each answer, so a confident wrong answer costs more than a hesitant one. The company also says both models give the best score for their cost: no model scores higher than d1.2 for less money, and no model scores higher than d1.2-ultra at any price. On calibration, it reports that across 13 benchmarks and 22,505 questions answered by all five models, d1.2-ultra’s calibration error is 0.8 points, versus 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. An internal table also lists accuracy on receipt value reading, document personal-data extraction, RefCOCO localization, COCO counting and multi-select, and ESC-50 / FLEURS audio questions; those figures are company results on held-out splits of public datasets.
Both models are available today on the Liquid API as d1.2 and d1.2-ultra, with a key from console.liquid.ai, at $0.02 and $0.042 per million input tokens respectively, with input-only billing. Liquid’s example: a 1024×1024 image costs 1.5 tokens per 32×32-pixel patch, or 1,536 tokens; a request bills its text and pictures once, and each extra question adds only its own text. The company says the models are coming to OpenRouter and Vercel soon, without a firm date.
Key points:
- Official introduction: https://x.com/liquidai/status/2122346077231923698 ; official blog: https://www.liquid.ai/blog/d1-2 ;
- These are concrete new model names in Liquid AI’s decision-model line, not the d1-ultra already stored on 17 October 2026, and not the earlier term LiquidAI d1-3B. OpenRouter and Vercel availability are still described as coming soon — do not write them as already shipped.
Advertisement
Explain
Why it matters
This is a new model name Liquid AI confirmed on 16 November 2026 with an official account post and a blog post, not a routine version bump of the existing term d1-ultra. Around 15 October 2026, official materials introduced d1-ultra as “the World’s Best Decision Model,” and that term was stored on 17 October. Inside this 00:32 scan window, the official introduction post (15:08:30Z / 23:08:30 Beijing time) and blog present d1.2 and d1.2-ultra as new versions with audio input and five new tasks, and give them their own index scores and API list prices, so they are filed separately.
This page only restates the official introduction post and the official blog. The Decision Index score of 75, the “best score for their cost” claims, the calibration-error comparisons, and the internal task accuracies are company statements; this page does not independently reproduce them.
Limits
OpenRouter and Vercel availability are still described as coming soon — do not write them as already shipped. The five new tasks do not yet have a public benchmark; Liquid says it evaluated them on internal test sets built from held-out splits of public datasets. Do not treat this page as a guarantee of extraction quality, calibration, or price, and do not describe these models as a replacement for the existing terms d1-ultra or LiquidAI d1-3B.
Usage
- Read the official introduction at https://x.com/liquidai/status/2122346077231923698 and cross-check the official blog at https://www.liquid.ai/blog/d1-2 ;
- Follow the official instructions to request a key at console.liquid.ai and call the Liquid API with the model names d1.2 or d1.2-ultra; wait for a further official notice before assuming OpenRouter or Vercel are live;
- When citing the Decision Index score of 75, the per-million-input-token list prices, the calibration-error figures, or the internal five-task accuracies, attribute them to the 16 November 2026 official post or blog, and recheck the upstream source;
- Do not describe d1.2 / d1.2-ultra as the d1-ultra already stored on 17 October 2026, or as the earlier term LiquidAI d1-3B; this page only restates official materials — recheck the upstream sources before citing them externally.