APIUp to 28% cheaper than official pricesTry the API →
Hermes
Back to News

LiquidAI's d1.2-Ultra Tops Decision Index With Near-Perfect Confidence Calibration

Liquid AI released d1.2 and d1.2-ultra , omnimodal decision models with text, image, and audio input. d1.2-ultra tops the Artificial Analysis Decision Index v2.1 at 75, two points ahead of prior best.

LiquidAI's d1.2-Ultra Tops Decision Index With Near-Perfect Confidence Calibration
Source
Liquid AI
Published
Author
Liquid AI
Read
4 min read

Takeaways

  • Liquid AI released d1.2 and d1.2-ultra , omnimodal decision models with text, image, and audio input.
  • d1.2-ultra tops the Artificial Analysis Decision Index v2.1 at 75, two points ahead of prior best.
  • Five new answer types: value read, span extraction, bounding boxes, number with range, multi-select.
  • Calibration error of 0.8 for ultra and 1.2 for d1.2, well ahead of competitors.
  • Pricing: $0.02 and $0.042 per million input tokens, input-only billing, available on Liquid API .
  • 98% receipt read accuracy, 94% RefCOCO localization, 95 to 97% on audio classification tasks.

Liquid AI has released d1.2 and d1.2-ultra, two decision models that accept text, images, or audio and return schema-constrained answers with probabilities. On version 2.1 of the independent Decision Index , d1.2-ultra scored 75, the highest listed result and two points above the previous leader. The release post also introduces five answer types, audio input, and input-only usage pricing.

Typed fields and calibrated probabilities make these models useful in automated extraction and classification pipelines. Applications can validate the output schema, accept high-confidence results, and send uncertain cases to a person or another system.

Five answer shapes, one endpoint

Each API request supplies source material and one or more questions. The service returns a typed result for every named question, allowing a single call to extract an invoice date, locate an object, identify sensitive text, and estimate a quantity.

TaskReturned dataExample
Value extractionText, number, or date, with a not found fallbackRead the total due from an invoice
Passage detectionMatching passages and their positionsFind every piece of personal information
Image localizationOne or more bounding boxesLocate every damaged component
Numeric estimationAn estimate and rangeCount objects in a photograph
Multi-select classificationEvery applicable option with an individual probabilityIdentify all objects present in a scene

The API accepts a recording as the material being evaluated or as the question itself. A caller can ask which language is spoken, whether a siren is audible, or submit the instruction as recorded speech instead of text.

Confidence becomes a routing signal

Calibration measures how closely confidence scores match observed accuracy. Liquid reports a calibration error of 0.8 points for d1.2-ultra across 22,505 questions from 13 benchmarks, with all five compared models answering the same questions. Lower values indicate closer agreement between confidence and accuracy.

ModelCalibration error
d1.2-ultra0.8
d1.21.2
Jev 1.142.7
Perplexity Decider v1.33.7
GPT-6 Luna on the Decisions API7.2

In a well-calibrated batch, answers assigned 80% confidence should be correct about 80 times out of 100. That relationship supports operational thresholds, such as accepting results above 95% and routing the rest for review. Calibration does not establish whether the underlying accuracy is sufficient, so teams still need task-specific testing.

Internal tests cover the new ground

No public benchmark yet covers all five new answer types, according to Liquid. The company assembled held-out internal test sets from public datasets and reported the following results.

TaskDatasetd1.2-ultrad1.2
Receipt value extractionCORD, SROIE98%96%
Personal-data detectionNemotron PII94% precision, 93% recall92% precision, 88% recall
Object localizationRefCOCO family94%94%
Object count within oneCOCO93%92%
Multi-label object presenceCOCO87%90%

Audio results ranged from 95% to 97% on ESC-50 sound identification and FLEURS language identification across 43 languages. Those evaluations used three to six answer options per question, making them narrower than open-ended audio understanding.

Pricing favors batched questions

Liquid lists d1.2 at $0.02 per million input tokens and d1.2-ultra at $0.042 per million input tokens. Neither model carries an output-token charge.

ModelInput priceOutput price
d1.2$0.02 per million tokens$0
d1.2-ultra$0.042 per million tokens$0

Image inputs consume 1.5 tokens for each 32-by-32-pixel patch at the applicable text rate. A 1024-by-1024 image therefore uses 1,536 input tokens. Multiple questions in the same request share that image input, while each question adds its own text tokens. Liquid says OpenRouter and Vercel integrations are planned.

The request is a question map

The endpoint accepts a source in the state field and a map of typed questions. Authentication uses a bearer token, and one request can collect several fields from the same input.

Strong fits and exposed edges

Document processing, personal-data redaction, visual inspection, intake triage, and audio tagging align with the API's typed outputs. A single decision request may replace combinations of OCR, regular expressions, classifiers, and custom confidence heuristics. General-purpose models remain better suited to drafting, conversation, and open-ended generation.

Published benchmark figures leave several production questions unresolved, including latency, throughput, domain shift, poor scan quality, noisy audio, and calibration on a customer's own data. The d1.2 model scored 90% on multi-label object presence, three points higher than d1.2-ultra's 87%. Liquid does not explain that reversal, so exhaustive-label workloads warrant testing on both models.

Self-hosting remains a promise

Liquid's models join decision-focused offerings from Perplexity, OpenAI, and Jev on the Artificial Analysis index. Their combination of constrained outputs, low input prices, and reported calibration below one point gives developers another option for extraction and classification systems that require explicit abstention and review thresholds.

Liquid says it plans to release the model weights, but it has not announced a date, license, or deployment requirements. Until those details arrive, d1.2 and d1.2-ultra remain hosted API products, and any future self-hosting case depends on the terms of that release.

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Sources

  1. 01x.com
  2. 02liquid.ai
  3. 03artificialanalysis.ai
  4. 04console.liquid.ai