Liquid d1.2 is $0.02. Ultra is $0.042
On November 16, 2026, Liquid AI priced d1.2 at $0.02 and d1.2-ultra at $0.042 per million input tokens, with no output charge.

On November 16, 2026, Liquid AI priced d1.2 at $0.02 and d1.2-ultra at $0.042 per million input tokens, with no output charge.

On November 16, 2026, Liquid AI published d1.2: State-of-the-Art Omnimodal Decision Models ↗. @liquidai ↗ posted it at 15:08 UTC. The picture above is ours. We have not called either model.
d1.2 and d1.2-ultra take text, images, and audio. Audio can be the recording you ask about, or the question, spoken instead of typed. Five new tasks: read a value and return the characters, or "not found"; return every matching span and its position; return a box, or a box per instance plus a count; estimate a number and a range; pick every option that applies, each with a probability.
The call is POST https://api.liquid.ai/decisions/v1/systemone, model d1.2 or d1.2-ultra, with a key from console.liquid.ai ↗ under Dashboard, then API Keys. The blog says OpenRouter and Vercel are coming soon.
The decision-model docs ↗, fetched the same day, still list the hosted pair as d1.1 and d1-ultra. The make-decisions reference ↗ still documents choice, score, and noul, with a 65,536-token limit for d1 and d1.1 and 32,768 for d1-ultra. It does not name d1.2. Liquid's invoice sample uses "type": "number" and prints None when the invoice states no VAT. We did not run it.
Both models bill input tokens only. Per million input tokens, the blog lists d1.2 at $0.02 and d1.2-ultra at $0.042. An image costs 1.5 tokens per 32×32-pixel patch, so a 1024×1024 image is 1,536 tokens. Text and pictures are billed once. Each extra question adds only its own text.
Artificial Analysis ↗, at 15:26 UTC, printed its own measurements:
| Model | Decision Index v2.1 | Place | $ per 1M index tasks |
|---|---|---|---|
| d1.2-ultra | 75 | #1 | $133 |
| d1.2 | 69 | #6 | $63 |
AA says 75 is a 4-point gain over d1-ultra and 69 is a 3-point gain over d1.1. On AA-Paperclip, AA scores ultra at 76, 3 points ahead of Drex Max v1.1. AA lists a 64K context window and text-plus-image input. Liquid's blog does not print 64K, and Liquid's blog adds audio. Liquid's own index line is 75, two points ahead of a previous best it does not name, with no higher score below d1.2's price and none above ultra at any price. The index page ↗ uses six benchmarks and penalizes a confident wrong answer.
Every number here is Liquid's, from the November 16 post.
Calibration error, on 13 benchmarks and 22,505 questions answered by all five models:
| Model | Calibration error |
|---|---|
| d1.2-ultra | 0.8 points |
| d1.2 | 1.2 |
| Jev 1.14 | 2.7 |
| Perplexity Decider v1.3 | 3.7 |
| GPT-6 Luna, Decisions API | 7.2 |
Liquid's reading of that gap: a stated 80% should be right about 80% of the time, so a low score can go to a person.
Liquid says no public benchmark covers the five new tasks. On its held-out splits: receipt reading (CORD, SROIE) 98% ultra and 96% d1.2; personal-data precision/recall (Nemotron PII) 94/93 and 92/88; object location (RefCOCO family) 94% both; a count within one (COCO) 93% and 92%; every object present (COCO) 87% ultra and 90% d1.2. On held-out audio, Liquid says both answer 95% to 97% of ESC-50 and FLEURS questions, 43 languages, three to six options or yes/no.
"Our new d1.2 (non-ultra) is not far behind at less than half the per-token price of the ultra." — Mathias Lechner ↗ (@mlech26l), November 16 at 15:25 UTC. His profile says "Cofounder/CTO at Liquid AI." The blog prices are $0.02 and $0.042.
"For classification and routing steps, the decision layer becomes a rounding error." — Mohammad Vazin ↗ (@mohammadvazin), November 16. The profile describes graphic design and a personal account.
"liquid's d1.2 isn't a chatbot, and that's the whole point." — Omar ↗ (@Om_Dream1), November 16. The profile says software engineering student.
"GLiDE v1.2 scores 73 and costs $780. Gemini Decision 1 sits at 66 for $345, GPT-6 Luna (Decisions) at 60 for $339." — Franco Naumow ↗ (@naumowf), November 16. His "d1.2 scores 69 for $63" matches AA's text. AA's text does not print the other three rows. Gemini 66/$345 and Luna 60/$339 match the rounded figures in our November 12 note (65.98 at $345, 60.25 at $339), an older snapshot.
Reddit was not reachable for this story.
Call d1.2 when the program wants a field, a label, or a probability, and a missing value should come back as "not found." Receipts, routing, moderation, and a threshold that sends the unsure rows to a person are the jobs the blog describes. A chat model in JSON mode writes the object, and those output tokens are billed. d1.2 returns the characters it read, or "not found," and Liquid does not bill output. That is the product difference. It is not a test we ran.
Budget $0.02 and $0.042, and label the 0.8. AA's cost for a million of its index tasks is $63 and $133. Decider v1.3, in our November 15 note, is also $0.02 per million input tokens, output free, with locate. Decision 1 is $0.10 on Google's November 11 post. The calibration table above is Liquid's. The decision-model docs say to calibrate any threshold on your own data.
Do not plan a d1.2 download. The November 16 post ships an API and does not link a weight file for these two ids. The same day's docs still send self-hosting to d1-2B, d1-omni-2B, and d1-omni-600M, under the LFM Open License, and they do not call those files d1.2. OpenRouter is the blog's "soon," not a route you can call today. AA's 64K line is AA's. For a checkpoint with a public card, start with the Decider note. Chat plans are in ChatGPT vs Claude vs Gemini. The harness around a routing call is the agent frameworks comparison.
Who should try it: teams pulling fields, routing, or gating review on the Liquid API, who will score their own files first. Who should skip it: anyone who needs these weights this week, an OpenRouter id today, or a paragraph back.