DEV Community

IA Compa
IA Compa

Posted on Originally published at iacompa.com

Liquid AI releases d1.2 and d1.2-ultra: decision models that add audio input and five new typed tasks

As of November 16, 2026, d1.2-ultra ranks No. 1 of 24 decision models on the Artificial Analysis Decision Index v2.1, at 74.8. Liquid AI's new d1.2 decision models take text, images and audio, return calibrated probabilities with zero output tokens, and bill input only, from $0.02 per 1 million tokens.

What changed with d1.2

Liquid AI released d1.2 and d1.2-ultra on November 16, 2026, as new versions of its hosted decision models on the Liquid API.

A decision model does not generate text: you give it a state and typed questions, and in a single pass it returns calibrated probabilities and structured answers, with zero output tokens. Both models now take audio input, so a question can be spoken.

The release adds five typed tasks: read a value from a document and cite its characters, or answer not found; find every passage of a kind (for example, personal data) with its position; locate objects in a picture with boxes and count them; estimate a number with a range; and pick every option that applies, each with its own probability.

On benchmarks, d1.2-ultra scores 74.8 on the Decision Index v2.1, ranking No. 1 of 24, at $133 per 1 million tasks and 218 ms median time per task; d1.2 scores 68.7. General LLMs score higher on the same index — Claude Haiku 5.5 (Max) at 77 and GPT-6 Luna (Max) at 76 — at far higher cost per task. Pricing is input only: $0.02 per 1 million tokens for d1.2 and $0.042 per 1 million for d1.2-ultra, with images billed at 1.5 tokens per 32×32 patch. The models are coming to OpenRouter and Vercel soon, and d1.2-ultra has a 0.8-point calibration error across 13 benchmarks in Liquid's own tests.

d1.2 by the numbers

  • Decision Index v2.1, d1.2-ultra: 74.8 (Nov 16, 2026)
  • Decision Index v2.1, d1.2: 68.7 (Nov 16, 2026)
  • Input token price, d1.2 ($ per million tokens): $0.02 (Nov 16, 2026)
  • Input token price, d1.2-ultra ($ per million tokens): $0.042 (Nov 16, 2026)
  • Cost per 1 million tasks, d1.2-ultra: $133 (Nov 16, 2026)
  • Median time per task, d1.2-ultra: 218 ms (Nov 16, 2026)

Why d1.2 matters for business

Classify, route and extract steps in automations can run on a purpose-built decision model at $0.02–$0.042 per million input tokens and millisecond latency instead of a full generative model, and a calibrated probability lets you auto-approve confident cases and route the rest to a person. The new span, box and audio tasks extend that to scanned documents, photos and call recordings.

Sources

Top comments (0)