Liquid AI released d1.2 and d1.2-ultra on November 16, 2026. Artificial Analysis ranks d1.2-ultra first on its Decision Index v2.1 with 75 points, while d1.2 places sixth with 69. Source Source
The day job is less glamorous than the ranking: reading invoice totals, finding signatures and sorting support messages. Liquid says both models process text, images and, newly, audio, with five added API task types. Source
Five new API tasks
According to Liquid, the additions let developers ask the models to:
- Extract a value: Return text, a number or a date, along with the characters it was read from—for example, the amount due on an invoice.
- Find matching passages: Identify text spans such as names, phone numbers and email addresses, including their positions in a document.
- Locate objects: Return a bounding box around a signature, or boxes around every instance of an object, plus a count.
- Estimate a number: Supply an estimate and a range, such as a crowd count of 12 with a range of 10–14.
- Select every applicable option: Assign multiple labels to a support message, such as billing and delivery, with a probability for each. Source
For extraction, Liquid says the models are designed to answer “not found” when an invoice doesn’t contain the requested value. Inventing an extra line item would be a questionable contribution to accounting. Source
Liquid says audio can be either the material being analyzed or the question itself. Its examples include identifying a spoken language, detecting a siren in the background and accepting a recorded question instead of a typed one. Source
Benchmark gains and internal tests
On Decision Index v2.1, Artificial Analysis reports a four-point gain for d1.2-ultra over d1-ultra and a three-point gain for d1.2 over d1.1. Source
Liquid says no public benchmark measures the five new tasks yet; its reported results for those additions come from internal tests using held-out splits of public datasets. Source
Launch prices and access
Liquid says both models are available through its API, priced at $0.02 per million input tokens for d1.2 and $0.042 for d1.2-ultra. Both bill for input tokens only, and the company says the new answer types use the same endpoint and request format. Artificial Analysis also lists first-party API availability and the same token prices. Source Source
Separately, Artificial Analysis reports costs of $63 per million Decision Index tasks for d1.2 and $133 for ultra. Those figures describe its benchmark workload—not a million tokens, and not a quote for your document queue. Source
OpenRouter and Vercel distribution are listed as coming soon. Liquid hasn’t supplied a firm date for either in the announcement. Source
Comments