Liquid AI's d1.2 Hits OpenRouter With 98% Extraction and Free Output Tokens
Liquid AI's omnimodal decision models d1.2 and d1.2-ultra hit OpenRouter, offering calibrated probabilities with input-only billing from $0.02 per million tokens.
Takeaways
- Liquid AI's d1.2-ultra and d1.2 decision models are now live on OpenRouter.
- Pricing is input-only: $0.02 and $0.042 per million tokens, no output charges.
- d1.2-ultra tops the Artificial Analysis Decision Index v2.1 at 75, two points ahead of prior best.
- Omnimodal inputs (text, image, audio) with 65,536-token context and calibrated confidence outputs.
- Five new tasks: value read, span extraction, bounding boxes, number-with-range, multi-select.
- Open weights promised eventually; see the Liquid blog post for details.
Liquid AI has added d1.2-ultra and d1.2 to OpenRouter following their launch on Liquid’s API. Developers can now access the models through OpenRouter using the IDs liquid/d1.2-ultra and liquid/d1.2.
The d1.2 family targets classification, extraction, localization, and routing workloads. Each model accepts a bounded question and returns a typed answer with a probability estimate, eliminating much of the prompt enforcement and output parsing required when a general-purpose language model generates free-form text.
One provider, no output-token fee
OpenRouter currently serves both models through Liquid as the sole provider. Requests go directly to Liquid, with no alternate-provider routing or fallback. The d1.2-ultra listing specifies a 65,536-token context window and support for text and image inputs.
Liquid documents audio support for the d1.2 family, although OpenRouter’s current description lists only text and images. Teams planning to send audio through OpenRouter should confirm endpoint support before building around that modality.
| Model | Input price | Output price |
|---|---|---|
liquid/d1.2 |
$0.02 per million tokens | No charge |
liquid/d1.2-ultra |
$0.042 per million tokens | No charge |
Images are billed at 1.5 input tokens per 32 by 32 pixel patch. Structured responses carry no output-token charge because billing covers the input only.
Teams already using an OpenRouter-compatible client can set the base URL to https://openrouter.ai/api/v1, supply an OpenRouter API key, and select either model ID. The single-provider setup makes availability and latency dependent on Liquid’s endpoint.
Confidence becomes a routing signal
A decision request defines a constrained answer type, such as a category, extracted span, number, or bounding box. The model returns a value and an associated probability that applications can use for thresholds, escalation, or automated review.
Liquid reports calibration errors of 0.8 points for d1.2-ultra and 1.2 points for d1.2 across 13 benchmarks containing 22,505 questions. Lower calibration error means the reported probabilities more closely track observed accuracy. Among comparable predictions scored near 0.7, for example, a well-calibrated model should be correct about 70% of the time.
Production thresholds still require validation against the deployment workload. Calibration can shift when documents, languages, image quality, or class distributions differ from the benchmark data.
Five schemas, plus abstention
The d1.2 release expands the available answer schemas beyond basic classification:
| Output type | Typical use |
|---|---|
| Value read | Transcribe a requested value from a document or image |
| Span extraction | Return the source text supporting an answer |
| Bounding boxes | Locate an object described in natural language |
| Number with range | Return a numerical result with an associated interval |
| Multi-select | Choose multiple valid options from a defined set |
An explicit “not found” result allows the model to abstain when the requested item is absent. Applications should still validate extracted values because abstention support does not guarantee that every unsupported answer will be rejected.
An invoice pipeline could ask for the amount due and receive the extracted value, its supporting characters, and a probability. A visual inspection system could describe an object and receive its image coordinates.
Accuracy, latency, and cost
At the time of Liquid’s release, the Artificial Analysis leaderboard listed d1.2-ultra at 75, two points above the previous leader, and d1.2 at 69. Liquid also published the following task results in its release notes:
| Task | d1.2-ultra | d1.2 |
|---|---|---|
| Receipt value extraction on CORD and SROIE | 98% | 96% |
| PII detection precision and recall on Nemotron PII | 94% / 93% | 92% / 88% |
| Object localization on the RefCOCO family | 94% | 94% |
| Object counting within one object on COCO | 93% | 92% |
On Liquid’s Decision Index workload, d1.2-ultra cost $133 per million tasks and recorded a median latency of 218 milliseconds per task. Liquid says the new generation reduces average task time by 28% compared with its predecessor.
The $133 figure is a benchmark workload estimate under token-based pricing, rather than a fixed per-task rate. Actual cost depends on prompt length, document size, image resolution, and retry volume.
Best fits and hard limits
The strongest fit is a pipeline that ingests a document, image, or supported audio clip, asks a bounded question, and takes an action based on the answer and probability. Common examples include invoice processing, receipt extraction, PII detection, content moderation, visual question answering, object localization, and support-ticket routing.
Input-only billing can reduce costs for high-volume decision workloads, particularly when a generative model would otherwise produce and validate structured JSON. The final comparison depends on input size, accuracy targets, fallback rates, and the cost of reviewing low-confidence cases.
Workloads requiring generated prose, tool calls, open-ended answers, or extended multi-step reasoning remain better suited to generative models or agent systems. The d1.2 models are specialized components for bounded decisions and extraction.
A growing category with one open question
Google’s Gemini Decision 1 also targets classification through a dedicated decision API with no output-token billing. Liquid broadens the category with additional answer schemas and audio support through its own API, reflecting demand for models designed around typed production decisions.
Liquid has said that it will open the model weights “at some point,” without providing a date or licensing terms. Released weights could enable self-hosted extraction and classification pipelines; until then, OpenRouter offers a hosted access path for teams already using its API.