Liquid AI’s d1.2-ultra scores higher than d1-ultra on a decision-model benchmark, but the upgrade is not a clean win for every job. OrcaRouter’s comparison says the new model is better at document tasks and handling rephrased questions, while its predecessor does better on some knowledge and answer-withholding measures.
Watch Desk analysis
What happened
OrcaRouter compares the two Liquid AI models using Artificial Analysis’s Decision Index v2.1. It gives d1.2-ultra a score of 74.85, first among 24 models, against 70.58 and sixth place for d1-ultra. The comparison says both cost $0.042 per million input tokens, while d1.2-ultra has a 66,000-token context window to its predecessor’s 33,000 and takes less time per benchmark task.
The results are not uniformly in the new model’s favour. OrcaRouter reports that d1-ultra scored higher on its omniscience accuracy and non-hallucination measures, and on raw accuracy for the HLE multiple-choice test. These are the comparison’s reported benchmark results, not a guarantee of performance on a particular team’s data.
Key findings
- A stronger composite score
d1.2-ultra scores 74.85 against 70.58 for d1-ultra on Decision Index v2.1. - Better document results
OrcaRouter reports a higher document all-pass rate for d1.2-ultra, 67.7% against 51.3%. - Different strengths on uncertain questions
The comparison gives d1-ultra higher omniscience accuracy and non-hallucination scores. - Repeated questions can change the bill
OrcaRouter says the newer model bills attached text and images once per request, rather than charging for them again with each additional question.
Why it matters
A better overall score does not automatically make a model the better choice for every part of a workflow. Document extraction and decisions that can be checked against a confidence threshold may suit d1.2-ultra, while a task that depends on factual judgement or a reliable “I don’t know” may call for testing the older model too.
The repeated-question billing difference could matter for teams asking several questions about the same document or image. OrcaRouter’s worked example estimates a substantial saving, but the result depends on the number of questions and the input size. One question per document does not produce the same saving.
Our read
This is a useful reminder that model upgrades are not always a one-way ladder. The headline score points towards d1.2-ultra; the component results point towards choosing by task, not by podium position. If you use either model, test your own prompts and failure cases before swapping a production route. Benchmarks are a map, not the territory, even when the map has decimals.
What to watch
- Whether Liquid AI publishes comparable technical details for both models.
- How the reported differences hold up on independent evaluations and real workloads.
- Whether teams route document work and factual checks to different models.
Discussion spark: Would you use one model across a whole workflow for simplicity, or route document extraction and factual checks to different models based on these trade-offs?
Sources and evidence
- d1.2-ultra vs Liquid AI D1 Ultra: Same Price, 2x Context – OrcaRouter (16 November 2026, 17:23 UTC)
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.