Liquid AI shipped d1.2 and d1.2-ultra on November 16 — decision models that now accept audio alongside text and images, return bounding box coordinates for object location, and sit at the top of the Artificial Analysis Decision Index v2.1 at 75 points. The timing matters: Google, Microsoft, and Cloudflare all released competing decision models in 2026, and Liquid AI just outscored all of them on the benchmark that counts.
If you haven’t used a decision model before, the distinction from a language model is worth spelling out, because the use case is genuinely different.
What a Decision Model Actually Does
A language model generates tokens. A decision model answers questions about its input. You send it a state — a document, an image, an audio clip, or any combination — along with a list of typed questions. It returns typed answers with calibrated confidence probabilities. It bills zero output tokens, because there are no output tokens to bill.
That’s not a minor implementation detail. If you’ve been asking a general-purpose LLM to “extract fields from this invoice and return JSON,” you’re paying for the generation of that JSON, plus a parsing step, plus retries when the format breaks. With a decision model, you define the schema up front, the model answers directly, and you pay only for what you sent in.
At high volume — document processing pipelines, content moderation queues, agent routing layers — the cost difference is significant. d1.2 is priced at $0.02 per million input tokens. d1.2-ultra is $0.042. Output tokens: zero.
What’s New in d1.2
The first d1 models handled text. d1-ultra added images in October. d1.2 adds audio, completing the input modalities, and introduces five new answer types:
- Value read — extract a specific field value from a document
- Span extraction — return the exact substring from the input
- Bounding boxes — locate objects in an image, return rectangle coordinates
- Number with range — estimate a quantity with a confidence interval
- Multi-select — return multiple choices from a defined set
The bounding box answer type is the most interesting addition. You can ask “where is the signature on this scanned form?” and get a rectangle coordinate back — no vision LLM, no prompt engineering, no parsing. For document processing workflows that need to locate fields on structured forms, this is a meaningful simplification.
Benchmark: First Place
On the Artificial Analysis Decision Index v2.1, d1.2-ultra scores 75 — the highest among all decision models. The previous leaders were GLiDE v1.2 at 73 and Drex Max v1.1 at 72.6. Google’s Gemini Decision 1 scores 66. Microsoft Decision-1 scores 62. OpenAI’s Decisions API scores 60.
The calibration number matters as much as the index score. d1.2-ultra has a calibration error of 0.8 across 22,505 questions from 13 benchmarks. Calibration error measures whether confidence scores mean what they say — a model that says “90% confident” should be right about 90% of the time. At 0.8, the scores are reliable enough to build threshold-based routing on: auto-process when confidence is high, escalate to human review when it’s low.
The API
The endpoint is POST https://api.liquid.ai/decisions/v1/systemone. All d1 models use the same schema: a model, a state, and a questions object. Here’s an invoice extraction example from the official docs:
import os, requests
invoice = "ACME SA\nInvoice 2026-117\nTotal due: 1 234,50 EUR\nPay before 03/04/2026"
response = requests.post(
"https://api.liquid.ai/decisions/v1/systemone",
headers={"Authorization": f"Bearer {os.environ['LIQUID_API_KEY']}"},
json={
"model": "d1.2-ultra",
"state": invoice,
"questions": {
"payee": {"type": "text", "instructions": "Who issued the invoice?"},
"total": {"type": "number", "instructions": "How much is due?"},
"due": {"type": "date", "instructions": "By when must it be paid?"},
},
},
)
The response gives you typed values and probabilities for each field. No prompt to maintain, no JSON parser to babysit. Keys from console.liquid.ai. The models are also available on OpenRouter if you’re already routing there.
When to Use a Decision Model
The honest answer: for a surprisingly wide class of production problems. Routing incoming support tickets to the right queue, classifying documents for compliance, scoring content against a moderation policy, extracting structured fields from forms, grading LLM outputs in a CI pipeline — all of these have a defined answer space and don’t need prose generation. They’re better served by a decision model than by parsing an LLM’s output.
Use an LLM when you need open-ended reasoning, synthesis, or generation. Use a decision model when the answer space is bounded and volume is high. Most production pipelines have more of the latter than developers realize.
Decision models have gone from a niche concept to a recognized category in under a year. Google, Microsoft, Cloudflare, and TypeSafe all shipped entries in 2026. Liquid AI is currently leading the benchmark. At $0.02 per million input tokens with zero output billing, d1.2 is worth adding to your evaluation list.

