Returns typed yes/no, choice and score probabilities from text or images in one pass, with zero generated output tokens.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Zero-Token Decisions: Where d1 Replaces a Chat Model Call
Technical Specifications
What the model supports
Comparison
d1 or d1.1: Should You Stay on the Original Hosted Model?
d1.1 is described as the successor to d1 and is now the default hosted decision model in Liquid's docs. Both are hosted decision models billed on input tokens only, and Liquid's docs say every hosted model uses the same endpoint, question types and response format. Both accept text and images. The practical decision is whether to keep an integration pinned to d1 or move to the newer default. Liquid has not published a head-to-head benchmark of the two.
Choose d1 when you have already validated thresholds against it, route through a gateway that still lists d1, or need results comparable to your earlier tests. Choose d1.1 for new direct integrations, since Liquid's current examples and vision documentation use it as the default. Do not carry d1 thresholds over to d1.1 unchanged. Re-run your labeled holdout set, because Liquid has not published the probability behavior of either model.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Ticket and request routing
d1 is built for the choice, yes/no and score questions that sit at the front of support and operations pipelines. One request can ask whether a customer wants a refund, which team owns it and how urgent it is. Liquid reports text decisions in 200 to 300 ms, and typed answers remove JSON repair. Use it ahead of an LLM that drafts replies. Evaluate on labeled tickets first, since Liquid's results are self-reported.
Visual inspection and image screening
d1 accepts images alongside text. Liquid reports 85–97% accuracy sorting good and defective parts across four VisA production lines, without task-specific training. That makes it a candidate for camera-based checks, damage screening and screenshot classification where a probability is enough. The figures are vendor-reported demos, not independent tests, so measure defect recall on your own parts before relying on it.
Guardrails, moderation and approval gates
Probability outputs make threshold policies practical. Liquid's moderation example blocks above 0.8, allows below 0.2 and sends the middle band to human review. d1 can also act as an LLM-as-judge replacement or approve an agent's tool calls. Tune thresholds against your own labeled data. Because d1 can't justify a verdict in text, keep a human or generative reviewer for borderline cases.
Agent step selection and context compaction
Liquid demonstrates d1 choosing the next action in a web agent and deciding whether to keep, trim or drop each tool output in a coding agent's session. In that test it removed 52% of tokens while keeping every output the task needed. With screenshots, it solved 12 of 12 Wordle games. These are vendor demos. They point to d1 as a fast control layer beside a generative agent.