
d1.2-ultra vs Liquid AI D1 Ultra: Same Price, Twice the Context, and One Place It Got Worse
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 127 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 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 · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 148 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 123 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
Liquid AI d1.2-ultra costs exactly what Liquid AI D1 Ultra costs: $0.042 per million input tokens, no output tokens, the same endpoint, the same three question types you already send. It arrived on November 16, 2026, a month after the model it replaces, and it is better on the composite — 74.85 against 70.58 on Artificial Analysis Decision Index v2.1, first of 24 against sixth. So the upgrade is free and obvious, and you should take it. What makes the decision worth an article rather than a one-line changelog is that four of the sixteen published component metrics went backwards, and two of them are the ones a threshold-driven pipeline feels most.
The pattern is specific. The new model is much better at reading documents, much better under paraphrase, and slightly worse at knowing things and at declining to answer. If your pipeline uses Liquid AI D1 Ultra as an extraction engine, the successor is a straight win. If it uses it as a gatekeeper that is supposed to say "I don't know" when the case is ambiguous, the successor is a two-model decision, not a swap.
Both models on one instrument

This comparison is unusually clean because both rows were produced by the same organisation on the same version of the same board, which is not true of most upgrade decisions. Artificial Analysis ran d1-ultra and d1.2-ultra on Decision Index v2.1, so nothing here depends on comparing a vendor number to an independent one, or on an index revision that moved the goalposts between the two releases. The vendor's own headline for the new model — first place, two points clear — matches the board it names.
• Price — $0.042 per million input tokens for both. Identical, and unchanged by the upgrade.
• Decision Index v2.1 — d1.2-ultra 74.85 (#1 of 24) vs d1-ultra 70.58 (#6).
• Context window — 66k tokens vs 33k, per the board's technical specifications. Liquid's launch posts state no window for either model.
• Time per task — 218 ms vs 301 ms. The upgrade is also a third faster.
• Cost per million tasks — $133 for both, which is what you would expect at an identical token price and a similar token count per task.
• Brier score — 0.150 vs 0.176, lower being better. This is the single number that most directly says "better calibrated".
The billing rule changed, and it is worth more than the index points
Under d1-ultra, each question was billed as its own prompt including every attached image. Ask ten questions about one photograph and you paid for the photograph ten times, which for a 1024×1024 image means 1,536 tokens multiplied by ten before any text is counted. Under d1.2-ultra the request bills its text and pictures once, and each additional question adds only its own text.
Work it through on a realistic request: a 2,000-token document, one 1024×1024 image, ten questions of about 30 tokens each. The old rule charges roughly 35,700 tokens; the new one charges roughly 3,840. At an identical $0.042 per million that is $0.0015 against $0.00016 per request — a little over nine times cheaper for the same work, at the same published price, from the same vendor, with no announcement headline attached to it. At 100,000 requests a month, about $150 becomes about $16.
Both figures are arithmetic on the vendors' own billing descriptions rather than measurements of a live bill, and the ratio depends entirely on how many questions you ask per state. Ask one question per document and the two models cost the same. Ask ten, and the "free" upgrade is the largest cost change in the release.
What the new model actually does better

Four components moved up, and two of them moved a long way.
• AA-Paperclip, 30% of the index — 75.81 against 64.69 on the decision score, and 67.7% against 51.3% of documents fully correct. Eleven points of document all-pass rate on the component that carries the most weight is the whole story of a twelve-point composite gain. This is the number to look at if you extract fields from invoices, receipts or resumes, because a document that misses one of fifteen answers counts as a miss either way.
• AA-Symmetry, 15% — 86.34 against 78.66, with all-pass consistency up from 45.8% to 56.7%. Reword the question, move the options around, negate a yes/no, and the new model holds its answer far more often.
• When2Call, 10% — 96.49 against 93.83. Tool calling was already the family's strength; this is a smaller gain on a component that was near its ceiling.
• EvalSafe-O*NET-AA, 15% — 89.53 against 88.15. A point and a half on workplace-document judgement, which is close to noise and should not be sold as an improvement.
• AA-Omniscience-Decision v1.1, 15% — 60.80 against 64.39 on the decision score. Down four points on the component that is half knowledge and half honesty.
• HLE-MultiChoice v1.1, 15% — 46.36 against 47.39. Down a point on the component where the whole field is weak.
The two regressions, in plain terms
The omniscience drop is not a scoring artefact; it shows up in the underlying accuracy and abstention rows, and they disagree with each other in a way that is worth understanding before you migrate.
• AA-Omniscience-Decision accuracy — d1-ultra answered 28.0% of questions correctly, d1.2-ultra 23.6%. The predecessor knew more.
• AA-Omniscience-Decision non-hallucination — d1-ultra 88.6%, d1.2-ultra 83.8%. The predecessor was also better at not inventing an answer. Both rows moved the same direction, which means this is a genuine capability change rather than a shift in how cautiously the model spends its abstentions.
• HLE-MultiChoice accuracy — 6.09% against 2.36%, the new model last of the 24 on raw accuracy. Its non-hallucination rate on the same set rose, 56.9% to 76.1%, so the drop in accuracy is partly bought with more abstention. But not entirely: the omniscience pair above shows the model answering less AND getting less right.
The consequence for a pipeline is narrower than it sounds. If you threshold on probability and send low-confidence cases to a human, a model that abstains more is doing you a favour, and the 83.8% non-hallucination figure is still sixth of 24. If you use the model as an automatic gate on questions that require knowing a fact — is this claim consistent with policy, does this code match the spec — the cheaper predecessor is measurably better and there is no reason not to keep using it for that slice.
What you have to change in your client
Almost nothing, and that is the reason to do it now rather than in a quarter. The endpoint is the same, api.liquid.ai/decisions/v1/systemone, called with "model": "d1.2-ultra"; the three question types are unchanged; responses are still structured JSON with a probability per option and no output tokens. The migration is a string change and a redeploy.
What is new is additive, and it is worth knowing because it removes workarounds rather than adding features. Five answer types join the existing ones: read a value from the page and return the characters it was read from, or "not found"; return every span of a kind with positions, which covers PII redaction without a second model; return a box around one object or every instance plus a count; return an estimate with a range; and return every applicable option with its own probability. Audio input works in both directions — the recording can be the subject, or it can be the question, spoken rather than typed.
The accuracy claims for those five tasks are Liquid's own, on internal test sets built from held-out splits of public datasets, because no public benchmark covers them yet. On the vendor's numbers the two models are close: 98% against 96% reading a value from a receipt, 94%/93% against 92%/88% on PII spans, 94% each on object location, and on the last two tasks the older model is not the weaker one on every row — the new model wins counting within one at 93% to 92% and loses picking every object present at 87% to 90%.
Running the old one and the new one side by side
Keeping both is the honest migration, and it is cheap: they are different model strings on the same endpoint and both are on Liquid's own API today, at the same price, with keys from console.liquid.ai. Neither is on our catalogue, and this piece is not an availability claim for either.
Where a routing layer becomes useful here is that the split between the two models is a per-question decision rather than a per-service one. The successor is the one to send documents, images and anything you will threshold on; the predecessor is the one to send fact-heavy judgements where its accuracy and abstention rows are still better. Compose that as two routes behind one API — the routing DSL lets you split by question type or by state size rather than by hand-coded service boundaries — and the two models stop competing for the same slot. Nothing about our side changes the $0.042 either model costs: provider list prices are passed through at 0% markup, so a Liquid price move is live here the same day rather than at the next billing cycle. And when a vendor API has a bad afternoon, automatic failover is what keeps the document route alive instead of turning one provider into a single point of failure for the whole pipeline.
The upgrade, in one line each
Take d1.2-ultra for documents, for images, for ambiguous states, and for anything you threshold on — it is eleven points better on the document component, eight better on symmetry, a third faster, doubles the context window and, if you ask more than one question per state, cuts your bill by an order of magnitude.
Keep d1-ultra for fact-heavy gating, where its omniscience accuracy and non-hallucination rows are both still higher, and where the HLE accuracy is nearly three times the successor's, small as both numbers are.
Skip the migration entirely only if your states carry no images, one question each, and no threshold — the case where the two models cost the same and the new one is simply better on every evaluation you use. In that case the only reason to wait is the nineteen hours it takes to find out that nothing broke.

