Liquid AI’s d1-ultra evaluates text, images, or both and returns probabilities over defined outcomes without generating output tokens. Its launch price is $0.042 per million input tokens, with images charged at the same token rate as text.
Developers can access it directly through Liquid’s API for bounded decisions: which department should receive a support ticket, whether a photograph shows damage, or how urgently an incident needs attention.
The output contract distinguishes it from a generative LLM. Developers define possible outcomes and receive probabilities their software can act on, without asking the model to write a label or explain a judgment. Liquid also claims improved accuracy and calibration, although its benchmark superiority claims remain vendor-reported and are not independently established by the supplied evidence.
A Decision Model Answers Within Defined Boundaries
According to Liquid’s decision-model documentation, these models evaluate a situation and return structured probabilities in a single forward pass. They do not generate an answer token by token.
d1-ultra supports three question types:
- Noul: A yes/no question that returns a probability between zero and one.
- Choice: A selection among named categories, with a probability distribution across the options.
- Score: A position on an ordered scale, using a defined scoring rubric.
For ticket routing, a Choice question could distinguish billing, technical support, and account access. The same ticket could be assessed for sensitive personal information with a Noul question and rated for urgency with a Score question.
Multimodal input extends that pattern to decisions involving pictures. Liquid’s announcement gives examples including identifying a circuit-board defect against a specification and checking a damage claim against a customer’s photograph. These proposed applications do not establish that the model meets a particular factory’s inspection standard or an insurer’s claims requirements.
A generative LLM remains the appropriate tool for an explanation, summary, conversation, or piece of code. d1-ultra is intended for answers within a defined set of outcomes. It could help decide where a customer message goes without also being responsible for composing the reply.
Why Calibration Matters Beyond Getting the Label Right
Accuracy measures whether the selected answer is correct. Calibration asks whether the probabilities attached to those answers are trustworthy.
Among comparable cases assigned an 80% probability of a positive outcome, for example, roughly 80% should actually be positive if the probabilities are well calibrated. This is a statistical property across cases, not a guarantee about an individual prediction.
Those probabilities can guide automation. A routing system might automatically handle cases with strong evidence for one category and send ambiguous ones to a person. In visual inspection, uncertain images could be escalated instead of forcing every item into a pass-or-fail decision.
Liquid says it specifically trained d1-ultra to produce more reliable probabilities because existing decision models can be overconfident. That is a useful objective, but developers still need to check calibration on their own inputs. Probabilities that behave well on an evaluation set may behave differently with unfamiliar product photos, unusual customer language, or changing class frequencies.
The threshold also depends on the cost of an error. Missing a dangerous defect and unnecessarily flagging a harmless item are not equivalent mistakes. A probability distribution gives an application more flexibility than a bare label; the application still needs an explicit policy for acting on it.
Input-Only Pricing Has a Per-Question Catch
The headline price is unusually small: $0.042 for one million input tokens, with no output-token charge. As an illustrative calculation, one million questions consuming 1,000 billed input tokens each would cost $42 at that rate.
Images use the same monetary rate, but Liquid first converts them into token usage. Its launch announcement specifies 1.5 tokens per 32 × 32-pixel patch, making a 1024 × 1024 image equivalent to 1,536 input tokens.
At the announced price, that image component costs:
- $0.000064512 for one image evaluated against one question.
- Approximately $64.51 for one million such image-question evaluations.
Those calculations exclude the accompanying text and question instructions.
Crucially, each question is charged as its own prompt, including its text and all supplied images. Asking several questions in one request therefore does not mean paying for the image only once.
A request asking three questions about the same 1024 × 1024 image would incur 4,608 billed tokens for the image component alone. Budgeting needs to account for each question and its associated context, not simply each API request.
Comparisons with an LLM workflow also need to include the full cost of actual prompts, number of questions, output requirements, retries, and escalation path. The input rate is attractive, but it does not by itself establish savings against every generative model.
The Direct API Is the Clearest Starting Point
Liquid’s launch announcement identifies its own API as available immediately and describes Vercel and OpenRouter access as coming soon. The broader documentation also lists those gateways, creating a discrepancy between the supplied sources. Direct Liquid API access is the unambiguous launch route; developers should verify gateway availability separately.
The hosted API documentation specifies an API key from the Liquid Console and the endpoint:
https://api.liquid.ai/decisions/v1/systemone
Requests select d1-ultra through the model field, provide context in state, and define named entries under questions. A minimal text-only request follows this format:
curl -s https://api.liquid.ai/decisions/v1/systemone \
-H "Authorization: Bearer $LIQUID_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "d1-ultra",
"state": "My order arrived with a cracked screen.",
"questions": {
"damage_report": {
"type": "noul",
"instructions": "Does this message report product damage?"
}
}
}'
The response places the probability under the named answer, in this case answers.damage_report.noul. The documented response format also reports input-token usage and zero output tokens.
For image requests, Liquid’s announcement demonstrates base64-encoded image data URLs in a top-level images array, alongside the textual state and questions. Developers can use that direct HTTP path to test multimodal decisions without first building a generative chat workflow.
Liquid’s Benchmark Lead Is a Claim to Test
Sources
- d1-ultraliquid.ai
- decision-model documentationdocs.liquid.ai
- hosted API documentationdocs.liquid.ai
