Liquid AI d1

Overview

d1 is the first decision model from Liquid AI, released on October 5, 2026, and it now handles both text and images. It works differently from the usual language model: it generates no tokens at all. It reads the unstructured data you give it together with your questions in a single forward pass and returns a probability for each possible answer. A text decision comes back in 200 to 300 ms, fast enough to sit in a real-time path.

It answers three kinds of questions: noul is a yes or no question, returned as a probability between 0 and 1; choice picks one label among many, with a probability per label; score gives a position on a scale, weighted by the probability of each level. One request can ask several questions about the same state, which saves input tokens.

The authors compared d1 against GPT-6.1 Sol and Claude Opus 5.5 on six real applications, from filtering support tickets to inspecting circuit boards. d1 matched or beat GPT-6.1 Sol on four of them, at 1/19 to 1/200 of the cost of both models, and answered faster on every task.

You can call it today by creating an API key at console.liquid.ai, where the model name is simply d1, or try it without writing code in the d1 Playground. It is also available through Vercel AI Gateway and OpenRouter, though both serve text only for now, with vision coming soon.

Key Features

• Decisions without generated tokens: It takes unstructured data plus one or more questions, reads them in a single forward pass, and returns probabilities directly without producing any text, with text decisions in 200 to 300 ms
• Three question types: noul answers yes or no as a probability between 0 and 1; choice selects one label among many with a probability per label; score returns a position on a scale weighted by the probability of each level
• Several questions per request: Multiple questions can be asked about the same state in one request, so the input tokens are paid for once, which suits making several judgments about one ticket or one image
• Text and image input: Images are passed as base64 data URLs in the images field; the visual inspection demo covers circuit boards, candles, cashews, and chewing gum across four production lines at 85% to 97% accuracy, and the model was never trained for those tasks
• Billed on input tokens only: There is no output token cost; images are counted as input tokens at the same rate, 1.5 tokens per 32x32 pixel patch, so a 1024x1024 image costs 1,536 tokens, and each question is billed as its own prompt including its text and all images
• Several ways to reach it: The model name on the Liquid AI API is simply d1, and the d1 Playground lets you try it directly; Vercel AI Gateway and OpenRouter also list it, though both are text only for now

Use Cases

• Teams that need to embed a judgment rather than generated content into a business process, such as deciding whether a ticket is a cancellation request or which folder a document belongs in
• Latency-sensitive real-time paths, where the authors report text decisions returning in 200 to 300 ms
• Teams trimming tool output inside a coding agent, where the demo removed 52% of tokens while keeping every output the task needed
• Industrial visual inspection and similar steps that require judging an image, reading it straight from a camera frame or screenshot instead of transcribing it into text first

Pros

• Generates no tokens and returns probabilities in one forward pass, with text decisions in 200 to 300 ms
• Matched or beat GPT-6.1 Sol on four of six applications in the authors' comparison, at 1/19 to 1/200 of the cost of both comparison models
• One request can carry several questions about the same state, so input tokens are reused
• Visual inspection scored 85% to 97% accuracy in the demo without any training on those specific tasks, understanding the job from a short description
• Billing counts input tokens only, with images converted at the same rate as text, so cost is predictable
• The authors published their methodology: one run per application per model, list prices without prompt-cache discounts, d1 costed at $0.04 per million input tokens, and they note that some test questions were written after d1's pipeline was set

Pricing

d1 is billed on input tokens only, with no output token charges. Images are counted as input tokens at the same rate: 1.5 tokens per 32x32 pixel patch, so a 1024x1024 image costs 1,536 tokens, and each question is billed as its own prompt including its text and all images in the request. The authors' cost comparison prices d1 at $0.04 per million input tokens. You need an API key created in the Dashboard at console.liquid.ai. Pricing through Vercel AI Gateway and OpenRouter is set by those platforms, and both are text only for now.

Summary

What makes d1 interesting is that it separates having a model make a judgment from having it generate anything. The usual approach wraps a decision in a request for a language model to emit JSON, which is slow, expensive, and carries a pile of tokens you never wanted. d1 returns probabilities directly, in one forward pass, in 200 to 300 ms, billed on input alone. The Context Compaction demo makes the value concrete: it reads each tool output in a coding agent session, decides whether to keep, trim, or drop it, removes 52% of the tokens, and keeps every output the task needed.

Its scope has clear edges: it answers structured judgments and does not generate content. Vision has just landed, and Vercel and OpenRouter still serve text only. The authors also note that each comparison ran once per application per model and that some test questions were written after d1's pipeline was set, so treat those numbers as an order of magnitude rather than a rigorous benchmark.

Version History

• d1 decision model released with vision support (2026-10-05): Liquid AI released d1, its first decision model, supporting text and image input. It generates no tokens and returns a probability for each answer in a single forward pass, with text decisions in 200 to 300 ms. It answers noul, choice, and score questions, and one request can carry several questions about the same state. Compared against GPT-6.1 Sol and Claude Opus 5.5 on six real applications, d1 matched or beat GPT-6.1 Sol on four, at 1/19 to 1/200 of the cost of both models, and answered faster on every task. Visual inspection covers circuit boards, candles, cashews, and chewing gum across four production lines at 85% to 97% accuracy without task-specific training. It is available through the Liquid AI API as model d1 and in the d1 Playground, with Vercel AI Gateway and OpenRouter live but text only for now. Billing covers input tokens only, with no output token charges