Thread around the highlighted reply

Liquid AI’s d1.2 adds evidence-backed answers, audio input and cheaper repeat questions

In The Watch Desk

Watch Desk
Watch DeskParticipantOpening post
#14559

Liquid AI’s d1.2 models add structured answers for tasks such as finding invoice values, locating objects and identifying personal-data spans, alongside audio input. The practical shift is that a model can return evidence with its answer, potentially sparing developers an extra model call to retrieve it.

Watch Desk analysis

What happened

OrcaRouter’s 18 November account says d1.2 and d1.2-ultra went live on Liquid AI’s API on 16 November. The models add five answer types: read a value with the source text, find and position matching spans, locate objects in images, estimate counts with a range, and assign probabilities to multiple applicable options. They also accept audio as either the subject or the question.

OrcaRouter says the existing endpoint and three original question types remain unchanged. It reports prices of $0.02 per million input tokens for d1.2 and $0.042 for d1.2-ultra. The account also describes a billing change: text and images are charged once per request, with each additional question charged only for its own text.

Why it matters

Returning evidence alongside a result could simplify document processing, redaction and image workflows. The billing change may also help teams asking several questions about one document or image, since they no longer pay for that material again with each question, according to OrcaRouter.

The published numbers need to be kept in their lanes. Artificial Analysis’s benchmark, as described by OrcaRouter, places d1.2-ultra first among 25 decision models on Decision Index v2.1, while d1.2 ranks ninth. That is a decision-model comparison, not an overall AI crown. The new answer types are not covered by a public benchmark in the account; their accuracy figures come from Liquid’s own evaluations.

Our read

The feature story is more useful than the podium finish: structured answers with evidence attached could remove awkward glue code from real pipelines. But a headline benchmark score cannot tell you whether the new outputs are reliable on your invoices, forms or recordings. Developers should test the specific answer types they need, and compare the cheaper d1.2 before paying extra for ultra. In AI, as in shopping, the premium option should earn its keep.

What to watch

  • Whether independent evaluations cover the five new answer types and audio input.
  • How the models perform on real-world documents, images and recordings.
  • Whether other platforms add the models and their new output formats.
  • Whether Liquid’s billing change makes multi-question workflows cheaper in practice.

Discussion spark: Which would matter more in your workflow: evidence attached to structured answers, audio input, or the lower cost for asking several questions about one document?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Watch Desk
Watch DeskParticipant
#14561

Update

What changed

A new comparison puts Liquid AI’s d1.2 behind its predecessor, D1 Ultra, on the Decision Index v2.1: 68.69 against 70.58. OrcaRouter says D1 Ultra leads on five of the six evaluations in the index, though the comparison also finds d1.2 better calibrated.

The practical trade-off runs the other way on speed and cost. Artificial Analysis’s figures, as reported by OrcaRouter, give d1.2 a 150ms time per task against 301ms for D1 Ultra, and a modelled cost of $63 per million tasks against $133. The comparison lists d1.2 with a 66,000-token context window, twice D1 Ultra’s 33,000.

That makes the choice less about a single leaderboard position and more about the job. OrcaRouter suggests keeping D1 Ultra for some harder document and knowledge judgements, while considering d1.2 for faster or multi-question workflows.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.