
d1.2 vs Liquid AI D1: Same $0.02, Six Components Won, and Two Accuracy Rows That Went Backwards
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- openaiNEWOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 80 tok/s
- openaiNEWOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 155 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 196 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 85 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 705 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 246 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 97 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 139 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 84 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 102 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 69 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 38 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 · 230 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 101 tok/s
Liquid AI d1.2 and Liquid AI D1 sit at the same list price — $0.02 per million input tokens, input billed only, no output tokens — on the same endpoint, answering the same three question types. They are two generations apart: D1 was the family's first decision model, which Artificial Analysis lists as released in September 2026, and d1.2 arrived on November 16, 2026, with d1.1 in between. Everything a comparison usually turns on — price, contract, client code, deployment shape — is identical. What is left is the model, and on the model the result is lopsided but not clean: d1.2 wins all six evaluations of the Decision Index v2.1 and leads by 7.9 points overall, while D1 still wins two of the raw accuracy rows underneath them.
Those two rows are the reason this is worth reading rather than skimming. Both sit on the knowledge axis, and both are the direct cost of the change that makes d1.2 better everywhere else — it declines far more often. The comparison is a template for what a "strictly better" upgrade looks like once you check the sub-rows: a big composite gain, a genuine behavioural improvement, and a small, predictable place where the new model gives something up.
The two rows, side by side

Both models were scored by Artificial Analysis on the same version of the same board, which removes the usual hazard of comparing a vendor number to an independent one. Neither figure here comes from Liquid.
• Decision Index v2.1 — 68.69 for d1.2 (#9 of 25) vs 60.78 for D1 (#21 of 25). Eleven models now separate them, and the gap is wide enough that no plausible rerun closes it.
• Price per million input tokens — $0.02 for both. Unchanged across two generations.
• Modelled cost per million tasks — $63 for both, which is what you would expect at an identical token price and a similar token count per task. Neither is cheaper; the board ranks both third.
• Time per task — 150 ms vs 151 ms. Effectively identical, and both faster than the 206 ms class median.
• Context window — 66k tokens for both, as listed by the board. Liquid's own posts state no window for either, so this is a third-party reading rather than a vendor specification.
• Brier score, lower being better — 0.164 vs 0.196. A 0.032 improvement in the probability attached to the chosen answer, at no cost.
The last row is the one to sit with. d1.2 is not just more often right; the confidence it attaches to its answer is closer to the truth, which is the property that makes a threshold pipeline work at all. And it arrives free.
Six components, and the size of each gap

The composite number hides a very uneven set of changes, and two of them are not really about accuracy.
• AA-Paperclip, document decisions — 62.60 for d1.2 against 54.92, with 44.3% of documents fully correct against 25.7%. That all-pass row is the biggest single improvement in the release for this pair: a fifteen-field invoice with fourteen right counts as a miss under either model, and d1.2 turns a quarter of documents into nearly half.
• When2Call, tool decisions — 96.09 against 81.59, with 95.5% accuracy against 77.9%. This is the one that changes what the model can be used for. D1 at 77.9% is a classifier you would put a human behind; d1.2 at 95.5% is one you would let route.
• AA-Symmetry, invariance — 76.52 against 58.67, with 50.0% all-pass against 32.5%. Reword the question, reorder the options, negate a yes/no: the old model's answer moved, the new one's mostly does not.
• EvalSafe-O*NET-AA, workplace judgement — 86.55 against 84.21, and 79.7% accuracy against 76.3%. The smallest gap on the board, and not a reason to migrate on its own.
• AA-Omniscience-Decision v1.1, knowledge — 60.94 against 58.31. Nearly three points, and the sub-rows underneath are the interesting part.
• HLE-MultiChoice v1.1, hard questions — 44.69 against 39.80, with 5.30% accuracy against 11.79%. The new model gets fewer hard questions right and still scores five points higher, because the index scores the probability rather than only the pick.
The abstention story is the real upgrade
Look at the non-hallucination rows together, because they describe a model that changed its character rather than its accuracy.
• AA-Omniscience-Decision non-hallucination — d1.2 66.6%, D1 3.3%. D1 answered 39.8% of these questions correctly and almost never declined; d1.2 answers 30.0% correctly and declines a great deal more.
• HLE-MultiChoice non-hallucination — 53.9% against 5.6%, on a benchmark where D1's raw accuracy is twice as high.
• AA-Symmetry all-pass — 50.0% against 32.5%, which is the same shift seen through consistency rather than through refusal.
D1 was the family's first model and it behaves like one: asked a question it cannot answer, it produces an answer. A 3.3% non-hallucination rate means that on roughly 97 of every 100 knowledge questions where the correct response was to abstain, it did not. If your pipeline sends low-confidence cases to a human reviewer, that is the property that decides whether the gate works; a model that never says "I don't know" gives you nothing to threshold on, and its 0.9 probabilities mean less than d1.2's 0.9s do.
The caveat is symmetric and worth stating: d1.2 gets fewer hard questions right. If your task is a knowledge question with a definite answer and no abstention path — is this claim consistent with the policy, does this code match the spec — you are trading accuracy for honesty, and the trade may not be free.
What the new model can answer that the old one cannot
d1.2 adds five answer types and audio input. D1 has none of them, and no amount of prompting closes that gap, because the response contract has no field for a span position or a bounding box.
• A value plus the characters it was read from, or "not found" — the extraction primitive that removes a second generative pass per field.
• Every span of a kind with positions — names, phone numbers, email addresses, as a redaction primitive rather than a summarisation prompt.
• A box around one object or around every instance plus a count.
• An estimate with a range attached.
• Every applicable option with its own probability, rather than one winner.
Audio works in both directions: the recording can be the subject, or it can be the question. The accuracy figures for all of this are Liquid's, on internal test sets built from held-out splits of public datasets, because no public benchmark covers the new tasks — 98% and 96% reading a value from a receipt for d1.2-ultra and d1.2, 94%/93% against 92%/88% on personal-data spans, 94% for both on object location. Treat them as the vendor evaluating itself; the board has no row for any of it.
One billing detail rides along. Under D1 each question was billed as its own prompt including every attached image; under d1.2 a request bills its text and pictures once and each extra question adds only its own text. Ask ten questions about one 1024×1024 photograph — 1,536 tokens of image — and the old rule charges about 15,360 image tokens; the new one charges 1,536. At an identical $0.02 the multi-question pattern is several times cheaper on d1.2, and the saving is in the billing rule rather than in the rate.
What breaks when you change the string

Less than you would expect, and the exceptions are worth a test rather than a worry. The endpoint stays api.liquid.ai/decisions/v1/systemone, the model string becomes d1.2, and the three original question types keep their request and response shape. A client that sends a value, a choice and a score gets the same JSON back.
Three things to check against your own sample before switching a production path:
• Parsing "not found". The new answer types can return an explicit absence, and a parser that assumed a non-null value will break on it. This is a feature, but it is a breaking change in a strict schema.
• Threshold placement. Probabilities mean something different now. A 0.8 on d1.2 is a more reliable 0.8 than a 0.8 on D1, so a threshold tuned on the old model's overconfidence will fire less often after the swap. Re-tune it rather than carrying it over.
• The knowledge slice. If part of your workload is fact judgement rather than document reading, measure that slice separately. It is the one place the new model's higher abstention could cost you coverage, and the board's accuracy rows say so plainly.
Who should migrate, and who should measure first
Migrate without ceremony if the workload is extraction, routing, tool selection or consistency-sensitive classification. Every published row favours d1.2, the price does not move, and the invariance and tool-calling gaps are large enough that a rerun is unlikely to reverse them.
Measure first if the workload is knowledge judgement at volume, or if the pipeline has a hand-tuned probability threshold. Neither of those is a reason to stay on D1 — but both are reasons to put a few thousand real states through both models and compare the probabilities rather than the labels.
Access first, since a migration question usually turns into a procurement question. D1 and d1.2 both live on Liquid's API; neither is on our catalogue, and nothing here is an availability claim. The model on our catalogue that belongs in this test is Jev 1.14, which takes EvalSafe-O*NET-AA to 91.14 where d1.2 stops at 86.55, speaks the same systemone shape, and bills $0.042 per million input tokens. Put all three on the same states and the routing DSL writes the split once — by question type, not by service — with provider list prices passed through at 0% markup so the $0.02 you see quoted is the $0.02 you pay. Automatic failover is what lets a model released two days ago hold one lane of a live workload instead of all of it.
That is the whole comparison. Same price, same contract, same endpoint, one generation of work between them, and six composite wins against two accuracy rows the old model still holds — both of which sit on the knowledge axis, where d1.2's abstention is what costs it. If your workload is extraction, routing or classification, every row you care about moved the right way. If it is fact judgement with no reviewer behind it, run the shadow comparison first and let the numbers on your own states decide.
