Liquid AI Ships d1.2 With Audio Input, No Independent Tests
Liquid AI's d1.2-ultra tops the Decision Index at 75, but its new audio input and five answer types ship with only internal benchmarks and no open weights date.
By Nathan Brooks
4 min read

What's News
- Liquid AI released d1.2 and d1.2-ultra on November 16, 2026 with audio input and five new task types
- d1.2-ultra scored 75 on Decision Index v2.1, the highest among decision models tested by Artificial Analysis
- Pricing stays at $0.02 and $0.042 per million input tokens; images cost 1.5 tokens per 32×32-pixel patch
- All five new tasks and audio carry only Liquid's internal benchmarks; no third-party evaluation exists
- Liquid reports calibration error of 0.8 points for d1.2-ultra across 22,505 questions, lowest among five compared models
Liquid AI released d1.2 and d1.2-ultra on November 16, 2026, adding audio input and five new answer types to its hosted decision-model API — and shipping every new capability with only the company's own internal tests behind it.
The launch consists of a company blog post with an invoice-processing code demo. The only independent evaluation, Artificial Analysis's Decision Index v2.1, covers the pre-existing benchmark suite, not the new features.
What actually shipped?
Both models accept text, images and audio through the Liquid API. Audio can be the material under evaluation — identifying a language or detecting a siren — or the question itself, spoken rather than typed. The API returns typed answers with probabilities, not free-form text.
Five new answer types arrived:
- Value extraction: reads a text, number or date from the input (an invoice total, say), returning the source characters or "not found"
- Passage detection: locates every passage of a specified kind, such as personal data, with its position
- Image localization: returns bounding boxes for one or all instances of an object
- Numeric estimation: provides an estimate with a range (12 people, range 10–14)
- Multi-select classification: returns every applicable label with its own probability
Both models carry a 64K-token context window. Pricing is input-only and unchanged from predecessors: $0.02 per million tokens for d1.2 and $0.042 for d1.2-ultra. Images cost 1.5 tokens per 32×32-pixel patch at the text rate. OpenRouter and Vercel integrations are said to be coming soon.
How did the models score independently?
Artificial Analysis, which runs the Decision Index independently, reported d1.2-ultra at 75 on version 2.1 — the highest score among decision models tested. Source [1] puts that four points above its predecessor d1-ultra at 71. But sources [2] and Liquid's own announcement [5] describe the lead as two points above the "previous best decision model" or "previous leader," suggesting a different baseline model at 73. The cheaper d1.2 scored 69, up three points from d1.1.
One caveat for practitioners: the index folds in the probabilities models assign to answers and penalizes confident mistakes more than uncertain ones. A score of 75 is not a 75% accuracy rate.
The index combines six evaluations spanning document decisions and extraction, tool calling, knowledge, reasoning and consistency under rewording. Its private AA-Paperclip component, weighted at 30%, tests 1,500 resumes, receipts, invoices and research papers; d1.2-ultra led there at 76, three points ahead of Drex Max v1.1.
Artificial Analysis estimated benchmark-weighted costs of $63 per million tasks for d1.2 and $133 for d1.2-ultra at current token prices.
What do the vendor's own numbers show?
Liquid AI states that no public benchmark measures the five new tasks yet. The company built held-out internal test sets from public datasets — CORD, SROIE, Nemotron PII, the RefCOCO family and COCO — and reported these results:
- Receipt value extraction (CORD, SROIE): 98% for d1.2-ultra, 96% for d1.2
- Personal-data detection precision/recall (Nemotron PII): 94%/93% vs. 92%/88%
- Object localization (RefCOCO): 94% for both
- Object count within one (COCO): 93% vs. 92%
- Multi-label object presence (COCO): 87% for ultra, 90% for d1.2 — a reversal the company doesn't explain
On audio, both models answered 95–97% of questions on ESC-50 sound identification and FLEURS language identification across 43 languages. Each question offered three to six options, including close confusions like Spanish vs. Portuguese, or yes/no questions — narrower than open-ended audio understanding.
Liquid also reports a calibration error of 0.8 points for d1.2-ultra across 13 benchmarks and 22,505 questions, versus 1.2 for d1.2, 2.7 for Jev 1.14, 3.7 for Perplexity Decider v1.3 and 7.2 for GPT-6 Luna on its Decisions API. These are Liquid's numbers on a test it chose; no outside group has published a comparison on the same questions.
What does this change for developers?
Liquid AI's decision models compete with Perplexity, OpenAI and Jev on the Artificial Analysis index. The combination of constrained outputs, low input prices and reported calibration below one point gives developers another option for extraction and classification pipelines that need explicit abstention and review thresholds.
What's still missing: no third-party evaluation of the five new tasks or audio capabilities, no latency or throughput data, no tests for domain shift, poor scan quality, noisy audio or calibration on customer data. On self-hosting, Liquid says "we will open the weights at some point" but has announced no date, license or deployment requirements.
Until independent benchmarks cover the new capabilities, buyers betting on audio or the five answer types are taking the vendor's word for it.
Source: epyhia.com
More from Nathan Brooks
Show full bio
News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.
2,230 articles