d1.2: State of the Art Omnimodal Decision Models | Blog
Liquid AI releases d1.2 and d1.2-ultra decision models with audio input and five new structured-output tasks. d1.2-ultra scores 75 on Artificial Analysis Decision Index v2.1, leading the field. Pricing at $0.02 and $0.042 per million input tokens respectively, with free output tokens. Models available now on Liquid API, coming to OpenRouter and Vercel soon.
Transcript
Onyx Liquid AI shipped d1.2 and d1.2-ultra today. Omnimodal decision models. Audio in, five new structured tasks, and ultra just took the top spot on the Artificial Analysis Decision Index v2.1.
Echo And it's a real release. Not a preview, no waitlist. On their API right now.
Onyx With aggressive pricing. Two cents per million input tokens for d1.2, four point two for ultra. Output tokens are free.
Echo Input-only billing, same pattern as their earlier decision models. Images cost 1.5 tokens per 32-by-32 patch at the text rate, so a 1024-square image is about fifteen hundred tokens.
Onyx Which for documents is basically nothing. Okay, the five tasks. One, read a value from text, like an invoice total. It returns the exact characters it read, or "not found." Not a guess. "Not found."
Echo That alone kills a whole class of hallucinations. Two, find every span of a type, personal data in a document, each with its position. Three, locate objects in images with bounding boxes, single or multiple.
Onyx Four, estimate a count with a range. Five, multi-choice, pick every label that applies, each with its own probability.
Echo And audio input works two ways. The context can be audio, like "what language is this" or "is there a siren." And the question itself can be spoken instead of typed.
Onyx So, the leaderboard. Decision Index v2.1, six benchmarks, decisions and extraction on resumes, receipts, research papers, tool calling, hard knowledge. Ultra scores 75. Two points ahead of the previous best.
Echo Two points is fine. The calibration error is what stopped me. Zero point eight, for ultra.
Onyx And everyone else is miles back. Jev 1.14 at 2.7. Perplexity Decider, 3.7. GPT-6 Luna on the Decisions API, 7.2.
Echo Seven point two. Ouch.
Onyx Because calibration error is the gap between predicted probability and actual correctness. If a model says 80 percent confidence —
Echo — it should be right 80 percent of the time. And at 0.8 error, you can actually trust the scores for routing. Low confidence goes to human review, the rest auto-approves.
Onyx Yeah, you got there first. That's the product story for me. Bounded decisions with confidence you can actually trust.
Echo On the new tasks, though… no public benchmarks yet. These accuracy numbers are internal test sets, held-out splits of public datasets. CORD and SROIE for receipts, Nemotron PII for personal data, RefCOCO variants for localization, COCO for counting and multi-label.
Onyx mm.
Onyx Still, the numbers. Ninety-eight percent on receipt reading for ultra, ninety-six for base. Ninety-four precision, ninety-three recall on PII extraction. Ninety-four on RefCOCO localization.
Echo Ninety-three within-one on COCO counting. And eighty-seven on multi-label COCO for ultra, where base d1.2 hits NINETY.
Onyx Wait, the base model wins?
Echo On that one task, yeah. Audio evals too, ninety-five to ninety-seven percent on ESC-50 sound classification and FLEURS language ID across forty-three languages. Three to six options per question, including close confusions like Spanish versus Portuguese.
Onyx Spanish versus Portuguese is mean.
Echo It's a good test. But until there's a public benchmark on these five tasks, it's Liquid's word. Which isn't nothing, their earlier decision models held up. It's not independent replication either.
Onyx Fair enough. Shipping it is easy, at least. Same endpoint, decisions/v1/systemone, same request format. You just add questions with the new types. Text, number, date, spans, box, boxes, estimate, choice with multiple true.
Echo The blog example reads three values off an invoice, then asks for a VAT that isn't there. Comes back None for that one.
Onyx Instead of inventing one, which models love to do.
Echo And the old d1, d1.1, and d1-ultra are deprecated. No audio, no new tasks. Migration is just switching the model string.
Onyx Plus OpenRouter and Vercel coming soon. That's the distribution play, OpenRouter gets you fallback across providers, Vercel gets you edge deployment.
Echo Into a crowded field, too. Jev, Perplexity Decider, GPT-6 Luna's Decisions API, Sage2, GLiDE, Dots, Clef. Decision models are a real category now, not one vendor's thing.
Onyx yeah.
Onyx Which makes two cents per million input feel like a shot fired. For the model that beats everyone on calibration? That's a different cost structure for document processing pipelines.
Echo If the calibration holds on your documents. That's the test. Console's at console.liquid.ai, grab a key from the dashboard, it's live now.
Onyx Drop in a few invoices, see if the confidence scores actually map to correctness. That's the pilot.
Echo I'd run that this week, honestly.
Onyx Catch you on the next one, Echo.