When a wrong yes costs more than a slower answer, calibration is the spec that matters. Liquid AI built d1.2-ultra for those harder calls. Released on 16 November 2026, it is Liquid's most capable hosted decision model, and it reports a calibration error of 0.8 across 13 benchmarks and 22,505 questions. That is lower than any rival Liquid measured.
The comparison set, as Liquid published it: d1.2 at 1.2, Jev 1.14 at 2.7, Perplexity's Decider v1.3 at 3.7, and GPT-6 Luna through its decisions interface at 7.2. These are publisher-reported, and lower is better.
Ranking and price
Liquid reports a score of 75 on the Artificial Analysis Decision Index v2.1, first among decision models and two points clear of the previous best. It also claims nothing scores higher at any price. At $0.042 per million input tokens with free output, "any price" is not a large number to beat.
Indicative rates, as always. Confirm in your own Liquid console.
Where the extra cost goes
Most of Liquid's internal tests favour ultra, though not all of them:
- Receipt value reading: 98%, against 96% for d1.2
- PII detection precision and recall: 94% and 93%, against 92% and 88%
- Counting within one: 93% against 92%
- Object location: tied at 94%
- Picking every object present: 87%, below d1.2's 90%
- Sound and spoken-language identification: 95 to 97% for both
Read those PII numbers twice. A five-point recall gap on personal data is the difference between a redaction pipeline you can ship and one that needs a human pass.
Spec sheet
| Field |
Value |
| API ID |
d1.2-ultra |
| Inputs |
Text, images, audio |
| Output |
Typed answers with probabilities; no generated tokens |
| Per-question budget |
65,536 tokens |
| Questions per call |
up to 128 |
| Media |
up to 8 images, up to 8 audio clips (300 s total) |
| Price |
$0.042 per 1M input tokens |
It shares every task type with d1.2: value reading, spans, bounding boxes, ranged estimates and multi-select. It replaces d1-ultra from October, which is now deprecated and lacks audio.
Open weights? Liquid has not committed to a date, and nothing in its documentation promises them, so plan on the hosted API.
Deciding whether ultra pays for itself
Split your traffic by risk. Run the routine slice on d1.2 and only the expensive-mistake slice, such as compliance flags, PII or payments, on ultra. Track false negatives on that slice for a fortnight. If ultra catches even a handful more, it has paid for the doubled token rate many times over. Compare other options under structured decisions.