Liquid AI releases d1.2 models, says ultra leads decision index


Liquid AI says its decision models add audio input and five task types. It says d1.2-ultra scored 75 to lead the Artificial Analysis Decision Index v2.1.
Sole source: Liquid AI says it has released d1.2 and d1.2-ultra through its Liquid API, adding audio input and five task types to its earlier decision models. An API lets developers use the models in their own software.
In its release announcement, the company says d1.2-ultra leads the Artificial Analysis Decision Index v2.1 with 75 points, two points over the previous best decision model. Liquid AI is the sole source for the release and its performance figures.
Liquid AI says a recording can be the material a developer asks about—for example, whether a siren is audible or which language a person speaks. A question can also be spoken rather than typed.
The company illustrates the five new task types with examples:

The interim final rule requires licenses to reexport advanced superintelligence (SI) chips to Indonesia, Malaysia, Singapore, Thailand or Brazil.
The company says no public benchmark yet measures the five new tasks, so it tested them internally on reserved portions of public datasets. In its reading and object-location tests, an answer counted as correct if it found the target or returned “not found” when the target was absent.
| Task and dataset | d1.2-ultra | d1.2 |
|---|---|---|
| Read a receipt value (CORD, SROIE) | 98% | 96% |
| Find personal data, precision / recall (Nemotron PII) | 94% / 93% | 92% / 88% |
| Locate an object (RefCOCO, RefCOCO+, RefCOCOg) | 94% | 94% |
| Estimate object count within one (COCO) | 93% | 92% |
| Select every object present (COCO) | 87% | 90% |
For personal-data extraction, precision measures how many returned passages were correct; recall measures how many of the relevant passages were found. On reserved sound and language recordings from ESC-50 and FLEURS, respectively,, Liquid AI says both models answered 95% to 97% of questions correctly. The questions were yes-or-no or offered three to six choices; the language dataset covered 43 languages.
Liquid AI also measured calibration error, the gap between a model’s predicted probability and how often it is right. Across 13 benchmarks and 22,505 questions answered by all five models it compared, the company says d1.2-ultra had an error of 0.8 points and d1.2 had 1.2, Jev 1.14 had 2.7, Perplexity’s Decider v1.3 had 3.7 and GPT-6 Luna on its Decisions API had 7.2. It says reliable probabilities can help developers send low-confidence tasks to human review.
Liquid AI charges only for input tokens, units used to measure material sent to a model; it does not charge for output tokens. It lists d1.2 at $0.02 and d1.2-ultra at $0.042 per million input tokens.
| Model | Price per million input tokens |
|---|---|
| d1.2 | $0.02 |
| d1.2-ultra | $0.042 |
Developers can obtain an API key from console.liquid.ai under Dashboard > API Keys. The new answer types use the same endpoint and request format as before. Liquid AI says the models are coming to OpenRouter and Vercel soon, without giving launch dates.
1 min

Attackers accessed driver’s license images and other personal information through a service provider, Uber says. The company is offering affected people free credit monitoring and fraud assistance.
2 min

Ofcom’s package also expands detection of known abuse material and requires some services to find previously unknown content; the regulator expects the duties to take effect in early 2027 if Parliament approves them.
2 min