Liquid AI launches d1-ultra, an API-only multimodal decision model: #1 on the Artificial Analysis Decision Index (65), ahead of Jev 1.14, at $0.042 per million input tokens

Why it matters

【Key points】

Liquid AI released d1-ultra on Oct 15 at 17:00 UTC and calls it its strongest decision model. Decision models do not generate text: they read the input in one pass and return a probability for each possible answer, which suits automated checks. d1-ultra reads text and images together, supports 44 languages and answers three question types: yes/no (a probability from 0 to 1), choice (a probability per label) and score (a position on a scale).

Liquid's numbers: 65 and first place on the Artificial Analysis Decision Index, at least 8 points ahead of other decision models; first in all eight categories of its own 112-evaluation suite, beating TypeSafe's Jev 1.14 throughout; and the lowest calibration error of any model it tested ("when it says 90%, it's right at least 90% of the time"). Example uses are spotting circuit-board defects against a spec and checking damage claims against customer photos.

d1-ultra is API-only, with no open weights, at $0.042 per million input tokens and no output-token charge; Vercel and OpenRouter are coming. Liquid says only that it is built on "a larger open-weight model" with heavy customization, without naming the base.

【Verifiable】

Blog: https://www.liquid.ai/blog/d1-ultra

Launch post: https://x.com/liquidai/status/2110777964489285875

【Our read】

1. Decision models got crowded in the past ten days: Liquid's Open d1 (e/723), OpenAI's Decisions API (e/717) and TypeSafe's $870M round for Jev (e/783). d1-ultra moves Liquid from open weights to a closed premium tier aimed squarely at Jev.

2. Calibration is the point: with trustworthy probabilities you can auto-approve above a threshold and route the rest to people, which matters more than raw accuracy.

3. The evidence is Liquid's own suite plus one Artificial Analysis index, and the base model is undisclosed; independent tests on real workloads are still needed.