Liquid AI's d1 decision model adds vision on the paid API

On 5 October 2026 Liquid AI published Introducing d1: The most capable decision model, now with vision. The post calls d1 Liquid’s first decision model and says it now takes text, images, or both. It also says that with last week’s experimental release, d1 became the first model to rival Jev on text decisions, and that extending those decisions to images is the new step. “Adds vision” matches that account. A separate Liquid post for the earlier text release did not turn up in a search of the company site. OpenRouter’s D1 page, opened the same day, lists a release date of 1 October 2026 and still describes the input as text.

What a call returns

Liquid says a decision model does not write tokens. d1 reads the input and one or more questions in a single forward pass and returns a probability for each allowed answer. A text decision, Liquid says, takes 200 to 300 milliseconds. OpenRouter’s provider table lists a 0.28 second P50 for the text route it shows. That figure is OpenRouter’s, for a page that still describes text input.

The post and the decision-model docs name three question types. Noul is a yes or no, returned as a probability between 0 and 1. Choice is one label among several, with a probability per label. Score is a position on an ordered scale, weighted by the probability of each level. Several questions can ride on one input. The docs say each question is still billed for its own text and for every image in the request.

Demos Liquid reports

The post says every demo runs live in the d1 Playground. These results are Liquid’s, from that playground, not an independent eval.

Visual inspection uses the public VisA set: circuit boards, candles, cashews, and chewing gum. Liquid says d1 separates good and defective parts at 85 to 97 percent accuracy, and that the model was never trained for the task and instead follows a short description. The methodology note says each comparison model also sees a known-good part from the same line next to the part under inspection.

Five text demos follow. A SQL filter asks yes or no of 150 support tickets. Code search walks the Hugging Face transformers tree, 6,511 files in Liquid’s count, one folder at a time. Smart folders files a document, then a search, into a folder and a subfolder. A web agent picks the next action on a flight-search page from a one-sentence goal. Context compaction keeps, trims, or drops tool output. Liquid says that demo removes 52 percent of the tokens and keeps every output the task needs. Four of the five, Liquid says, were adapted from pg-jev, jevgrep, jev-ultrafast, and fast-jev-compaction. It does not say which one was not.

On games, Liquid says a Tetris screenshot lifts the score from 70 to 81 cleared lines, that screenshot Wordle was solved 12 of 12 times at 3.8 guesses on average, and that Quick, Draw! recognition was 5.2 of 6 drawings among 62 words, against 0.6 for a random guess.

The six-app comparison

Liquid says it tested d1 against GPT-6.1 Sol and Claude Opus 5.5 on six real applications, from filtering support tickets to inspecting circuit boards. It says d1 matches or beats GPT-6.1 Sol on four of the six, costs 19x to 200x less than both models, and answers faster on every task. The post does not, in the text, name the four wins or the two losses.

Liquid’s methodology note says each application was run once per model on 5 October 2026. The chat models got one message each, answered in JSON, and used their default reasoning setting. Costs are list prices with no prompt-cache discount, and d1 is priced in that note at $0.04 per million input tokens. Smart Folders is 105 passages, with cost stated per 1,000. Liquid says it wrote six of the 15 code questions, and two of the four compaction sessions, after the d1 pipeline was set. No third-party scorecard is attached.

Where the model is served

Liquid says d1 is on its API, with keys at console.liquid.ai. The documented call is POST https://api.liquid.ai/decisions/v1/systemone. Billing is input tokens only, and output tokens stay at zero. Images are 1.5 tokens per 32 by 32 patch, so a 1024 by 1024 image is 1,536 tokens. Each question is billed again for its text and for every image. The docs say two questions on that image are 3,072 image tokens plus text, that remote image URLs are refused, and that the body must carry base64 (JPEG, PNG, WebP, or GIF).

The paid id d1 is the vision model. The docs say d1:free is text-only, and that the TypeSafe client cannot send images yet. The post says the Vercel and OpenRouter copies are text-only for now, with vision “coming to both soon.” OpenRouter’s page still describes text input, at $0.04 per million input tokens and $0 output, with a 65,536-token context shown as 66K.

Practical takeaway

Image input, on Liquid’s docs, is the paid d1 model on Liquid’s API. OpenRouter and d1:free are still text. The latency, the VisA range, the compaction cut, and the 19x to 200x claim are Liquid’s single pass on 5 October.