
Liquid AI d1.2 Ships With Audio In and Boxes Out — and the Cheap Model Is the Interesting One
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 180 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 223 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 98 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 79 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 199 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 114 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Liquid AI d1.2 and Liquid AI d1.2-ultra went live on the vendor's API on November 16, 2026, and the headline number belongs to the expensive one: d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, first among the 25 decision models the board ranks. The more consequential change is in the request body, not the leaderboard. Both models now return answers the previous generation could not express — a value plus the characters it was read from, a list of spans with their positions, a bounding box, a count with a range around it, a set of options each carrying its own probability — and both accept audio, either as the thing being asked about or as the question itself.
That is a different kind of release from an accuracy bump. A decision model used to hand back a label and a probability, which meant any pipeline that needed the evidence behind the label had to bolt on a second model to find it. d1.2 answers with the evidence attached, in one forward pass, at $0.02 per million input tokens. d1.2-ultra does the same at $0.042. Neither generates output tokens, so there is still nothing to stream: these are models you call and get a structured object back from.
Both are live today on Liquid's own API and nowhere else that the vendor will confirm. Everything below separates what Liquid reports from what Artificial Analysis measured, because on this release the two are not the same set of numbers and the difference matters.
What actually shipped, item by item
The five new answer types are additive. The endpoint is unchanged, the three original question types are unchanged, and a client that sends the same JSON it sent yesterday gets the same shape of response.
• Read a value — ask how much is due on an invoice and get the number plus the characters it was read from, or the string "not found" when the field is absent. This is the one that removes the most glue code, because the usual workaround is a second generative call to re-read the page.
• Find every span of a kind — every name, phone number and email address in a note, each with its position in the text. That is a redaction primitive, and it is the first time this family has offered one.
• Locate things in a picture — one box around the signature on a scanned form, or a box around every instance of an object plus the count. "Not found" is an available answer, which matters more than it sounds: an object detector that must always return a box will invent one.
• Estimate a number with a range — twelve people in the photo, plausibly ten to fourteen. The range is the honest part of the output, and it is the piece most teams would otherwise have to calibrate themselves.
• Pick every option that applies — billing and delivery, each with its own probability, rather than a single winner. Multi-label classification was previously faked by asking the same question once per label and paying for the state each time.
Audio arrives in both directions. The recording can be the subject — which language is being spoken, whether a siren is in the background — or the question can be spoken rather than typed. Liquid reports 95% to 97% accuracy on held-out recordings across ESC-50 and FLEURS, with three to six options per question and close confusions such as Spanish against Portuguese in the mix.
One billing change rides along and is easy to miss. A request now bills its text and images once, and each additional question adds only its own text. Under the previous generation every question re-billed the full prompt including every attached image, so a ten-question interrogation of one photograph paid for the photograph ten times.
The board, and what the first place is first among

Artificial Analysis has both new models scored on Decision Index v2.1, and the numbers are not vendor-reported: the board ran them itself, on the same six evaluations, at current list prices. d1.2-ultra lands at 74.85, first of 25 decision models and 2.2 points clear of its nearest decision-model rival, Drex Max v1.1 at 72.62. d1.2 lands at 68.69, ninth of 25, above the class median of 66.
The qualifier worth carrying around is that "first" is scoped. The same board also scores two general language models on the same instrument, and both beat d1.2-ultra outright: Claude Haiku 5.5 (Max) at 76.49 and GPT-6 Luna (Max) at 75.67. Neither is a decision model, and the price of the comparison is severe — $6,218 and $1,548 per million tasks against $133 — but "number one on the Decision Index" without the word decision in it is not what the board shows.
What the board does corroborate, independently, is the calibration story. Decision scores on this index subtract the Brier score of the probability attached to the chosen answer, so a model that is right but wildly confident is punished. d1.2-ultra carries a Brier of 0.150, d1.2 0.164, and the model they replace, d1-ultra, 0.176. The cheaper new model is better calibrated than the older flagship despite scoring 1.9 points lower overall, which is the sort of split that only shows up when somebody else runs the harness.
Time and cost complete the picture and pull in opposite directions. d1.2 runs at 150 ms per task, eighth fastest on the board and just under the 206 ms median. d1.2-ultra runs at 218 ms, slower than the median and 15th. On cost per million tasks the board puts d1.2 at $63, third cheapest of the 25, and d1.2-ultra at $133, eleventh against a median of $137.
The new-task numbers, and who produced them

No public benchmark covers any of the five new answer types yet, so every accuracy figure for them comes from Liquid's own internal test sets, built from held-out splits of public datasets. They are a vendor's evaluation of its own model on data it chose, and they should be read as exactly that.
• Reading a value from a receipt, on CORD and SROIE — 98% for d1.2-ultra, 96% for d1.2. A read counts as correct when it finds the target or correctly says it is absent, which makes the not-found case part of the score rather than a free pass.
• Personal-data spans, on Nemotron PII — 94% precision and 93% recall for the ultra, 92% and 88% for d1.2. The recall gap of five points is the largest separation between the two models on any of the five tasks, and it is the one that matters for redaction, where a missed span is the failure that gets you fined.
• Object location, on the RefCOCO family — 94% for both. No separation at all.
• Counting within one, on COCO — 93% for the ultra, 92% for d1.2.
• Picking every object present, on COCO — 87% for the ultra against 90% for d1.2, the single row where the cheap model wins.
So the premium buys a clear gain on reading values and on span recall, and almost nothing on spatial tasks. If your use case is boxes, the $0.02 model is the one to test first.
Audio, and the spec sheet that has not caught up
Audio is the claim with the widest gap between the announcement and the independent record. Liquid's post describes audio input as a headline capability of both models; Artificial Analysis's technical specification for both still lists input modality as text and image, and the index has no audio evaluation, so nothing on the board tests the claim. That is not a contradiction — the board simply has not got there — but it does mean the audio capability is currently a vendor statement with no third-party check behind it.
The same caution applies to the five new answer types: the index measures none of them. A model can be first on the board and unmeasured on the thing you actually want to do with it.
The vendor's calibration table is worth reading for a different reason. Across 13 benchmarks and 22,505 questions answered by all five models, Liquid puts calibration error at 0.8 points for d1.2-ultra, 1.2 for d1.2, 2.7 for Jev 1.14, 3.7 for Perplexity's Decider v1.3, and 7.2 for GPT-6 Luna on its Decisions endpoint. Those are vendor-run numbers on a vendor-chosen question set. The board's Brier scores point the same direction on the same two models, which is the closest thing to a second opinion available this week.
Checking the cost sentence against the board

Liquid makes a claim that is unusually falsifiable: no model scores higher than d1.2 for less money, and no model scores higher than d1.2-ultra at any price. Both halves survive the board, with the same scoping caveat as before.
• The second half is close to trivially true among decision models — nothing else on the 25-model board scores above 74.85 at all. It fails only if the two general language models are admitted, and they cost 12 to 47 times as much per million tasks.
• The first half holds on the decision board: every decision model scoring above d1.2's 68.69 costs more per million tasks. The nearest cheap challenger is Perplexity's Decider v1.3 at 70.41 and $68, which is under two points of index for five dollars per million tasks — the closest the board comes to a counterexample, and still on the right side of Liquid's sentence.
• The cost figures are Artificial Analysis's own, computed at current list prices and averaged with the index weights. They are a modelled cost per task, not an invoice, and they move if a provider changes a price.
That last point is where a routing layer earns its place rather than advertising itself. This release changed Liquid's billing rule, halved the effective price of the cheap model against the old flagship, and did it two days ago. On our side, provider list prices are passed through at 0% markup, so a vendor cut lands here the same day rather than at the next contract renewal — the $0.02 and $0.042 in this article are the vendor's numbers, not a marked-up restatement of them.
What to do this week, if you already run a decision model
Nothing in this piece is an availability announcement: both new models are on Liquid's API, and neither has reached our catalogue yet. The reason to mention ours at all is that the release is a decision-model release, and the decision model we do serve is Jev 1.14 — the one row on the board that beats d1.2-ultra on a component, EvalSafe-O*NET-AA at 91.14 against 89.53, over the same systemone request shape at $0.042 per million input tokens. Whatever you conclude about the five new answer types, the security-and-workplace slice has an incumbent you can benchmark against today.
The move that makes sense before the third-party platforms Liquid promises actually appear is a shadow run. Send the states your pipeline already sees to both d1.2 models, compare the probabilities rather than the top-1 labels, and look specifically at the cases where the model answers 0.8 and is wrong. That tail is where the previous generation was weakest, and it is the one thing this release claims to have fixed that the index can only partly confirm.
Automatic failover is the other half of trying a two-day-old model without betting a production path on it: route a slice to d1.2-ultra, keep the incumbent behind it, and let a bad afternoon at one vendor degrade the answer rate rather than take the pipeline down. That is a cheaper way to learn whether 75 is real than waiting for somebody else's benchmark.
The open question is not whether these models shipped — they did, with prices and an endpoint — but whether the five new answer types hold up outside Liquid's own held-out splits. Nobody has published an independent number for any of them, and until somebody does, the honest position is that this release adds a genuinely useful output contract on evidence that is entirely the vendor's own.
