Liquid AI marks a milestone in non-generative AI with the d1.2 series, delivering unprecedented decision accuracy and multimodal audio-visual reasoning.

On November 16, 2026, Liquid AI shifted the paradigm of machine intelligence with the release of d1.2 and d1.2-ultra. While the industry has been locked in a generative arms race focused on long-form text production, Liquid AI has taken a different, more surgical approach: the decision model. Unlike traditional LLMs that predict the next token, the d1.2 series is designed to return probability-based decisions on closed questions, providing a level of precision and reliability previously unseen in frontier models.
This release represents a historical milestone for the 'decision model' class of AI. By moving away from the stochastic nature of text generation and toward deterministic, high-confidence decision outputs, Liquid AI is providing the missing link for autonomous agents and industrial-grade automation. The d1.2 series doesn't just 'chat' about a problem; it solves it with mathematical certainty.
The d1.2 architecture is built to handle complex, multi-step reasoning across diverse input modalities. A standout feature of this generation is the integration of native audio input capabilities. Both d1.2 and d1.2-ultra can process raw audio to identify sounds or transcribe and classify spoken languages with extreme accuracy. This makes them uniquely suited for real-world sensory integration tasks.
The models also introduce five specialized API tasks designed for high-stakes data extraction and spatial reasoning. These tasks move beyond simple prompting, allowing developers to interact with the model for structured data retrieval, object localization, and numerical estimation. Furthermore, the models are engineered with a 'refusal mechanism' to prevent hallucinations; if the input does not contain the necessary information to answer a question, the model refuses to guess, a critical requirement for enterprise reliability.
The performance metrics for d1.2-ultra are nothing short of industry-leading. On the Artificial Analysis Decision Index v2.1, d1.2-ultra scored a 75, making it the first model to break the 70-point barrier and placing it two points ahead of its nearest competitor. Perhaps more importantly, d1.2-ultra represents the pinnacle of the index; no other model currently scores higher at any price point.
Calibration is where the d1.2 series truly separates itself from the competition. Calibration error measures how well a model's predicted probabilities match actual outcomes. d1.2-ultra achieved a calibration error of just 0.8 points across 13 benchmarks and 22,505 questions. To put this in perspective, this significantly outperforms Jev 1.14 (2.7), Perplexity Decider v1.3 (3.7), and even the massive GPT-6 Luna (7.2). This level of calibration ensures that when the model provides a probability, developers can trust that probability for downstream logic.
Liquid AI has introduced a highly efficient, developer-friendly pricing model. Unlike generative models that charge for both input and output tokens, the d1.2 series uses an 'input-token billing only' structure. Since the models return structured decisions rather than long text strings, users are not penalized for the complexity of the model's reasoning, only for the data they provide.
This makes the d1.2 series exceptionally cost-effective for high-volume tasks like document processing or real-time audio monitoring. For instance, a 1024x1024 image is billed at only 1,536 tokens, allowing for massive scale without the exponential cost growth typical of vision-language models.
The d1.2 series is optimized for applications where accuracy and reliability are non-negotiable. In the realm of automated finance, the 98% accuracy on receipt extraction makes it an ideal engine for expense management and automated bookkeeping. For cybersecurity and compliance, the high precision/recall on PII detection allows for seamless, automated data masking and privacy auditing.
In the field of robotics and IoT, the multimodal capabilities—specifically object localization (94% on RefCOCO) and audio recognition—enable more intelligent edge devices. Developers can build agents that not only 'see' and 'hear' their environment but make high-confidence decisions about what they are perceiving, rather than just describing it.
Developers can access the d1.2 and d1.2-ultra models immediately via the Liquid API at console.liquid.ai. The API is designed for seamless integration into existing workflows, supporting structured outputs that can be directly consumed by programmatic logic. For those preferring managed platforms, Liquid AI has announced that the d1.2 series is coming soon to OpenRouter and Vercel.
To begin, simply sign up for a Liquid AI developer account, generate your API key, and start sending multimodal requests. Whether you are processing text, images, or audio, the d1.2 series provides a robust, low-latency foundation for the next generation of decision-centric AI applications.
API Pricing — Input: 0.02 USD per million input tokens (d1.2) — 0.042 USD (d1.2-ultra), input only / Output: 0.00 USD (Input-only billing) / Context: Input-token billing only; images cost 1.5 tokens per 32x32-pixel patch