Liquid AI launches d1: an API‑only decision model that returns calibrated probabilities with zero output tokens

Liquid AI announced on September 29, 2026 the public launch of d1, an API‑only decision model that the company brands as d1:free. Unlike traditional large language models that emit text tokens, d1 returns only calibrated probability values, and the service reports zero output tokens for each call. The announcement was covered by MarkTechPost and is documented in Liquid AI’s own decision‑model reference pages. The model is positioned as a complement to generative AI, offering a deterministic, token‑free interface for a range of classification‑style tasks.

How d1 Works

According to the Liquid AI documentation, d1 accepts a JSON payload that combines a description of the current state with one or more typed questions. Each question is mapped to a specific primitive, and the model returns a set of calibrated probabilities that answer the query without producing any textual token stream. The service records the field usage.output_tokens as zero for every request, which the company highlights as a way to avoid token‑based billing and to simplify downstream processing. The probability outputs are designed to be directly consumable by applications that need a numeric confidence score rather than free‑form text.

Liquid AI defines three core primitives that shape the way questions are interpreted. The first, called Noul, is a probabilistic yes/no primitive that returns a single probability indicating the likelihood of a true answer. The second, Choice, produces a probability distribution over a predefined list of named options, enabling multi‑class classification in a single call. The third primitive, Score, returns an ordered rubric of scores, each associated with a numeric weight, which is useful for ranking or grading tasks. By constraining inputs to these typed forms, d1 can deliver well‑calibrated outputs across diverse domains.

The public endpoint for d1 is a POST request to /decisions/v1/systemone, as listed in the official API reference. Calls must include an authentication token and the JSON body described earlier. The response payload contains a field named “probabilities” that holds the calibrated values for each primitive invoked, along with metadata such as request latency and the zero‑token usage metric. Liquid AI’s docs note that the endpoint is hosted on the company’s cloud infrastructure and is not available for on‑premise deployment.

Intended Use Cases

Liquid AI markets d1 for a suite of decision‑oriented applications. Classification tasks such as spam detection or sentiment analysis can be expressed as Noul or Choice queries, delivering a confidence score that downstream systems can act upon. Routing decisions—determining which service channel or workflow step to invoke—fit naturally into the Choice primitive. Scoring and grading, for example evaluating content quality or compliance, leverage the Score primitive’s ordered rubric. The model is also promoted for moderation pipelines, reranking of search results, and as an LLM‑as‑judge component that validates the outputs of generative models without generating additional text.

At the same time, Liquid AI advises customers to continue using generative LLMs for tasks that require natural language creation, such as drafting emails, summarizing documents, engaging in dialogue, writing code, or performing multi‑step reasoning. In the company’s view, d1 excels when the problem reduces to a binary or multi‑class decision with a need for calibrated confidence, while generative models remain the tool of choice for open‑ended content generation. This division of labor is intended to lower token costs and improve the interpretability of decision points within larger AI pipelines.

In practice, an e‑commerce service could send a POST request to /decisions/v1/systemone containing the order details as the state and a Choice question that lists possible fraud risk levels such as “low”, “medium”, and “high”. d1 would return a calibrated probability for each level, for example 0.92 for “low”, 0.07 for “medium”, and 0.01 for “high”. The platform can then route orders with a probability above a chosen threshold to a manual review queue, while automatically approving the rest. Because the response contains only numeric scores, the integration avoids parsing text and can be executed in milliseconds, fitting real‑time decision pipelines.

Limitations and Operational Considerations

Liquid AI makes clear that d1 is offered solely as a hosted service. The company does not provide a downloadable model, nor does it allow customers to fine‑tune or otherwise train the decision engine on private data. This restriction simplifies compliance and version control but also means that organizations cannot run the model on isolated infrastructure or customize its internal parameters. Access is limited to the API endpoint, and pricing is based on request volume rather than token consumption, reflecting the zero‑output‑token design.

The fact that usage.output_tokens is reported as zero does not imply that the service incurs no computational expense. Liquid AI’s internal monitoring still tracks CPU and GPU cycles, memory usage, and latency, all of which contribute to the cost structure behind the scenes. The zero‑token metric is primarily a billing abstraction that separates decision‑making workloads from the token‑based pricing model used for generative LLM calls. Users should therefore anticipate that high‑volume decision traffic will still generate measurable charges, even though no text tokens are emitted.

With d1 now publicly available, developers can integrate calibrated decision logic into applications without managing token budgets or parsing free‑form text. The launch signals a broader industry move toward specialized AI services that complement, rather than replace, large language models. Organizations that need fast, interpretable confidence scores for routing, moderation, or scoring can adopt d1 today, while continuing to rely on generative models for creative and reasoning tasks. In practice, the new option reshapes how AI pipelines allocate compute, budgeting decision calls separately from text generation and potentially lowering overall operational costs.