Liquid AI releases d1-ultra for text and image decisions without generating prose

Liquid AI released d1-ultra on October 15, 2026, giving developers a model that judges text and images without writing a response. Available through the Liquid API, it returns probabilities for predefined questions. The company is targeting harder jobs, including circuit-board inspections and photo-based damage claims, at $0.042 per million input tokens.

Liquid AI’s stated starting point was the gap between easy decisions and difficult real-world checks. Its examples pair evidence with a standard: spotting a circuit-board defect against a specification, or checking a customer’s damage claim against a photo. These are judgments about supplied material, rather than requests to compose a reply. The company positions d1-ultra specifically for that harder end of decision-making.

To build it, Liquid used a larger open-weight base model than the one behind d1, then applied its own extensive customizations and training pipeline. The company says the changes brought particular gains on knowledge-related tasks. It also says it trained d1-ultra to produce more reliable probabilities, addressing what it describes as a tendency for decision models to be overconfident.

Liquid compared the new model with d1 and Jev 1.14 across 112 evaluations. It claims d1-ultra beat Jev across all tested tasks, combining higher accuracy with better-calibrated probabilities at the same price. Separately, Liquid says d1-ultra leads other decision models on the Artificial Analysis Decision Index by at least eight points. Both performance comparisons are claims presented in Liquid’s launch post.

The interface keeps d1’s three question formats. Developers submit unstructured text, images, or both, alongside one or more questions. The model processes that input in a single forward pass—one run through the model—and returns probabilities without generating output tokens. Its response is therefore a structured assessment of the supplied questions, not a written explanation of its reasoning.

• Yes or no: a binary question receives a probability between zero and one.
• Multiple choice: developers supply labels, and the model returns a probability for each.
• Score: the answer represents a position on a scale, rather than a free-form written assessment.

Liquid’s launch example sends a circuit-board photograph, a short description of the production-line scene, and a yes-or-no question asking whether the board has a defect. The returned field contains the probability for that decision. Vercel’s model listing describes classification, routing and scoring as the intended uses, and lists a 32,768-token context window—the amount of input the model can work with.

Liquid bills images at the same token rate as text. Its image accounting uses 1.5 tokens for every 32-by-32-pixel patch, making a 1024-by-1024 image count as 1,536 input tokens. There is no output-token charge because the model does not generate output tokens.

The important billing detail is that each question counts as its own prompt, including its text and all images. Submitting several questions about the same photograph does not mean paying for that photograph just once. The listed token rate therefore needs to be read alongside the number of questions an application asks.

For access, Liquid’s launch post says its own API is available now and names Vercel and OpenRouter as coming soon. Vercel AI Gateway already has a d1-ultra listing, identifying Liquid AI as the provider and October 15 as the release date. OpenRouter access remains a planned distribution step in Liquid’s announcement.