Skip to content
VECTOR WIREAI INTELLIGENCE
UTC
Home

LiquidAI's d1.2-ultra tops Artificial Analysis Decision Index with audio, five new task types

d1.2-ultra ranks first on Artificial Analysis Decision Index v2.1 among decision models.

LiquidAI released d1.2 and d1.2-ultra on November 16, updating its omnimodal decision models with audio input support and five new structured task types1. d1.2-ultra ranks first among decision models on the Artificial Analysis Decision Index v2.12.

Both models are available immediately on the Liquid API, priced at $0.02 per million input tokens for d1.2 and $0.042 per million input tokens for d1.2-ultra. The models bill input tokens only. LiquidAI says d1.2 and d1.2-ultra are coming to OpenRouter and Vercel.

New task types and pricing mechanics

The five new tasks cover structured extraction and spatial grounding: reading a value from input (text, number, or date), finding every passage of a kind in a text (spans), and locating objects in images via bounding boxes. The new answer types use the same endpoint and request format as existing tasks. A request bills its text and images once, and each additional question adds only its own text. A 1,024×1,024 image costs 1,536 tokens, calculated at 1.5 tokens per 32×32-pixel patch at the text rate.

Benchmark results

On the Decision Index v2.1, no model scores higher than d1.2-ultra at any price, and no model scores higher than d1.2 for less money, according to LiquidAI. d1.2-ultra's calibration error is 0.8 points across 13 benchmarks and 22,505 questions answered by all five indexed models; d1.2's calibration error is 1.2 points on the same set.

On the new capabilities, LiquidAI reports internal benchmark results. d1.2-ultra scores 98% on reading a value from a receipt, versus 96% for d1.2. For finding every piece of personal data in a document, d1.2-ultra scores 94% precision and 93% recall; d1.2 scores 92% precision and 88% recall. Both models score 94% on locating an object described in a photo. d1.2-ultra scores 93% on estimating the number of objects in a photo within one; d1.2 scores 92%. On picking every object present in a photo, d1.2 scores 90% and d1.2-ultra scores 87%.

For audio, both models answer 95% to 97% of questions about which sound is heard or which language is spoken on held-out recordings.

ANALYSIS The pricing structure charges only input tokens and amortizes document context across multiple questions, a design that favors high-volume extraction workflows such as invoice processing and PII detection. The internal benchmarks cover LiquidAI's own evaluation sets.

Was this useful?

The day’s vector, in your inbox.

Vector Wire Daily: one dispatch each publishing day.