Liquid AI Releases d1.2 and d1.2-ultra: Omnimodal Decision Models That Top the AA Decision Index
Liquid AI just dropped its new decision‑model line, d1.2 and d1.2‑ultra, and the hype train is already chugging full steam. The headline claim is simple: d1.2‑ultra tops the Artificial Analysis (AA) Decision Index at 75, with d1.2 trailing at 69.
Liquid AI just dropped its new decision‑model line, d1.2 and d1.2‑ultra, and the hype train is already chugging full steam. As a founder who runs real hosting infrastructure day‑in‑day‑out, I’m more interested in what the numbers mean for a business that actually bills customers for API calls, not the glossy press release. The headline claim is simple: d1.2‑ultra tops the Artificial Analysis (AA) Decision Index at 75, with d1.2 trailing at 69. But the real story is in the pricing, the token model, and the practical limits of a “zero‑output‑token” decision engine.
What the d1.2 family actually does
Both models are hosted decision services that answer typed questions about text, images or audio in a single API call. Unlike generative LLMs, they never emit output tokens – they only return a calibrated probability per option and a structured answer type. That means you only pay for the input you send. The endpoint is https://api.liquid.ai/decisions/v1/systemone and the service is currently hosted only; there are no downloadable weights, no disclosed parameter count, and no latency figures in the public docs.
The new d1.2 line adds five typed answer formats: “read a value” (text, number, date), “find every span”, “locate objects” (box or boxes), “estimate a number with range”, and “pick every option that applies”. Audio can be part of the query, either as the subject (e.g., “what language is spoken?”) or as a spoken question. The API consumes 16‑bit PCM WAV and counts 10 tokens per second of audio. On held‑out ESC‑50 and FLEURS recordings, the models answer 95‑97% of audio questions correctly – a respectable figure but still a “held‑out” claim, not an independent benchmark.
Scoreboard bragging rights and why they matter
The AA Decision Index v2.1 aggregates six evaluations, with AA‑Paperclip carrying 30% of the weight and When2Call 10%; the remaining four each get 15%. d1.2‑ultra’s 75 puts it ahead of Drex Max v1.1, SubQ Snap 1.2 and GLiDE v1.2, all at 73. d1.2 sits at 69, two points below the deprecated d1‑ultra (71). The index penalises confident wrong answers, so a higher score implies better calibration and lower over‑confidence – crucial when you’re routing decisions to downstream automation.
Calibration error numbers back that up: d1.2‑ultra reports a 0.8‑point error across 13 benchmarks (22,505 questions), while d1.2 sits at 1.2. By contrast, Perplexity Decider v1.3 is at 3.7 and GPT‑6 Luna (Decisions API) at 7.2. Lower error means you can set tighter confidence thresholds before falling back to a human, which translates directly into cost savings on review labor.
Pricing mechanics – token‑based, not per‑call
The pricing is straightforward: $0.02 per 1 M input tokens for d1.2 and $0.042 per 1 M for d1.2‑ultra. Images are billed by patch: a 1024×1024 picture costs 1,536 tokens (1.5 tokens per 32×32 patch). Audio is 10 tokens per second. The token count for a request is the sum of text tokens, image patches, audio seconds, plus the tokens for each question (the UI lets you set a per‑question token estimate). Shared state (the document, image or audio payload) is billed once per request; each additional question only adds its own token cost.
Because output tokens are always zero, the cost model is essentially “pay for what you send”. For a typical workflow – say a 1 000‑token text document, one 1024×1024 image, and three short questions (50 tokens each) – the token calculator in the UI shows a total of roughly 2 200 tokens. At $0.02 per million, that request costs a fraction of a cent. Even the ultra model, at $0.042 per million, remains cheap for low‑volume use cases. New customers also get $2 in free credit on Liquid Console, enough for a handful of test calls.
Real‑world task performance – where the numbers line up
The source material lists performance on several public datasets. For “read a value” from receipts (CORD, SROIE), d1.2‑ultra hits 98% accuracy, d1.2 96%. Personal‑data extraction (Nemotron PII) sees ultra at 94% precision / 93% recall, d1.2 at 92% / 88%. Object localisation (RefCOCO family) is 94% for both. Counting objects within one (COCO) is 93% (ultra) vs 92% (d1.2). Multi‑label selection (COCO) actually favours d1.2 at 90% vs 87% for ultra.
These figures are all from Liquid’s internal tests on held‑out splits, not third‑party benchmarks. Still, they give a sense of where the models excel: structured extraction from documents and object detection in images are solid. The audio claim (95‑97% correct on ESC‑50 and FLEURS) is promising but unverified outside the vendor’s own labs.
Production constraints you need to know
The API caps each request at 65,536 input tokens and a 4.5 MB body. You can send up to 128 questions in a single call, which is generous for batch processing but can hit the token ceiling quickly if you’re feeding large images or long audio streams. The media limits are eight images, or exactly one image for box/boxes tasks, and up to eight audio clips totalling 300 seconds. Those limits are hard‑coded; you can’t simply stream a video feed into the model.
Because the service is hosted only, you’re dependent on Liquid’s uptime and latency. No latency numbers are published, so you’ll have to benchmark in‑house. For latency‑sensitive applications – say real‑time fraud detection – you may need to factor in network round‑trip time and the fact that the model runs a full forward pass for each request.
Business‑risk lens – should you adopt?
If your product hinges on cheap, high‑throughput decision‑making – for example, routing insurance claims, triaging support tickets, or extracting fields from invoices – the token‑only billing and low calibration error are attractive. The ultra model offers higher accuracy on tougher tasks at roughly double the per‑token price, but still under a cent per thousand tokens, which is negligible compared to human review costs.
However, the lack of open weights means you can’t self‑host for data‑sovereignty or cost optimisation at scale. If you’re building a platform that needs to guarantee service continuity regardless of a third‑party outage, you’ll have to keep a fallback model (perhaps an older d1‑ultra or an open‑source alternative) in the loop. Also, the pricing model assumes you can batch many questions into a single call; if you need per‑question latency, you’ll end up paying the higher ultra rate more often.
Actionable advice for founders and hosting providers
First, run a quick token audit on your existing pipelines. Use the UI calculator or the token formulas (text tokens + 1 536 per 1024×1024 image + 10 per second of audio) to estimate per‑call costs. If you’re under a few thousand tokens per request, the d1.2 pricing is already cheaper than most hyperscaler LLM APIs that charge for output tokens as well.
Second, prototype the new answer types. The “read a value” and “find every span” endpoints can replace brittle regex pipelines for invoice processing. The “locate objects” and “estimate a number with range” are handy for visual inspection tasks – think automated quality checks on product images.
Third, build a confidence‑threshold router. This will keep your automation pipeline lean while preserving accuracy.
Finally, plan for redundancy. Keep an older decision model or an open‑source alternative on standby, and monitor Liquid’s service health. The token‑only billing is a win, but vendor lock‑in remains a risk until the weights are released.
Bottom line: Liquid AI’s d1.2 family offers a cheap, well‑calibrated decision service that can handle text, images and audio in a single call. The ultra variant tops the AA Decision Index and delivers modest accuracy gains at about double the token price. For independent hosting providers and SaaS founders, the real value lies in the token‑only cost model and the new structured answer types – provided you manage the hosted‑only limitation and build a fallback strategy.
— Allan Ali, Founder
This article was produced with AI-assisted research and editorial support. Reporting is based on the source material cited below. Sources: MarkTechPost; marktechpost.com; Global1.News (19 November 2026).
By Allan Ali, Global1.News
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Wow
0
Sad
0
Angry
0
Comments (0)