Two cents per million input tokens, and nothing for output, because there is no output in the usual sense. Liquid AI's d1.2, released on 16 November 2026, answers typed questions about text, images and now audio with structured results and probabilities. It is the default model in Liquid's hosted decision line and the cheaper half of a pair; d1.2-ultra sits above it at roughly twice the price.
New this release
Audio is the headline. A clip can be the subject of a question (which language is this, is that a siren) or the question itself, spoken rather than typed. Alongside that come five new task shapes:
- Read a value as text, number or date, or report that it is not there
- Spans: find every passage of a given kind, with positions
- Box / boxes: locate one object or every instance in an image
- Estimate a number with a range
- Multi-select: choose every option that applies, each with its own probability
The older hosted models, d1, d1.1 and d1-ultra, are now deprecated. They still answer yes/no, choice and score questions but get neither audio nor the new tasks.
Limits to design around
| Limit |
Value |
| API ID |
d1.2 |
| Tokens per question |
65,536 (state, images and audio included) |
| Questions per call |
128 |
| Images |
up to 8 (exactly 1 for box questions); JPEG, PNG, WebP |
| Audio |
up to 8 clips, 300 seconds total; WAV, MP3, FLAC, OGG |
| Request body |
under 4.5 MB including Base64 media |
| Price |
$0.02 per 1M input tokens, output free |
Token accounting is predictable. An image costs 1.5 tokens per 32 x 32 patch, so 1024 x 1024 comes to 1,536 tokens. Audio costs 10 tokens a second. Shared state and media are billed once per request, and every extra question adds only its own text, which makes batching many questions against one document very cheap. Prices are indicative; confirm on your Liquid console.
Watch the 4.5 MB ceiling. Uncompressed WAV hits it at around 100 seconds of 16 kHz mono, well before the 300-second audio cap.
Picking d1.2 over ultra
Liquid reports d1.2 at 96% on its receipt value-reading test, 94% on object location and 90% on picking every object present, which is actually above the ultra model's 87% on that last test. Its calibration error across 13 benchmarks is 1.2, against 0.8 for ultra. Both sets of figures are publisher-reported, measured on held-out splits of public datasets.
Proving it on your data
Choose one decision your pipeline makes thousands of times a day. Send 500 labelled examples through d1.2, bucket the answers by stated probability, and check whether the 0.8 bucket is right about 80% of the time. If it is, you can automate everything above your threshold and send the rest to a person. Only move to ultra if the hard cases are where you lose money.
It is available on the Liquid API and listed on Vercel AI Gateway as liquid/d1.2, even though Liquid still described Vercel as coming soon at launch. Browse other structured decision models or the full Liquid AI catalogue.