Liquid AI adds image support to d1 decision model

Liquid AI has introduced d1, its first decision model, with support for both text and images. Rather than generate a written response, d1 processes unstructured input and one or more questions in a single pass, returning probabilities for possible answers. The company says text decisions take 200 to 300 milliseconds and can cover yes-or-no questions, selecting among labels, or assigning a score. A request can include several questions about the same input, which Liquid AI says can reduce token use.

Liquid AI says it evaluated d1 against GPT-6.1 Sol and Claude Opus 5.5 across six practical applications, including support-ticket filtering and circuit-board inspection. The company reports d1 matched or exceeded GPT-6.1 Sol on four of the six applications. Liquid AI also says d1 cost 19x to 200x less than both comparison models and responded faster on every task. Its published methodology says each model ran once on each application on October 5, 2026.

For visual inspection, the model sorted acceptable and defective items across circuit boards, candles, cashews, and chewing gum, with reported accuracy ranging from 85% to 97%. Liquid AI says d1 had not been trained for that inspection task and followed a short description. Its five text demonstrations include sorting 150 support tickets, searching a 6,511-file code repository, organizing documents, operating a flight-search website, and trimming outputs in a coding-agent session. In the context-compaction demonstration, Liquid AI reports d1 removed 52% of tokens while retaining the outputs needed for the task.

The examples illustrate the intended role of decision models: handling structured choices that Liquid AI says can otherwise require more expensive language-model calls. Builders can access d1 through the Liquid AI API, and the company says image input can be sent as base64 data URLs. Billing is based on input tokens, with no charge for output tokens; an image is counted at 1.5 tokens per 32×32-pixel patch, making a 1024×1024 image 1,536 tokens. Liquid AI says each question is billed as a separate prompt that includes its text and any images.