A generated title card headed ‘Half the price, twice as fast’, subheaded ‘Liquid AI d1.2 against Liquid AI D1 Ultra’. Two large rounded cards sit side by side: the left, headed ‘d1.2’, reads ‘$0.02’ over ‘per 1M input tokens’ plus ‘150 ms per task’; the right, headed ‘D1 Ultra’, reads ‘$0.042’ over ‘per 1M input tokens’ plus ‘301 ms per task’. A caption beneath reads ‘the older flagship still wins five of six components’, and three outlined chips read ‘66k vs 33k context’, ‘Brier 0.164 vs 0.176’ and ‘measured by Artificial Analysis’. The OrcaRouter mark is composited bottom right.
Guides & Insights

d1.2 vs Liquid AI D1 Ultra: Half the Price, Two Points Behind, and Better at the Job You Repeat

Author

Gideon Frost

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Liquid AI d1.2 costs $0.02 per million input tokens. Liquid AI D1 Ultra, the model that was the family flagship until November 16, 2026, costs $0.042. On the Artificial Analysis Decision Index v2.1 the older, dearer model is still ahead — 70.58 against 68.69 — and it is ahead on five of the six evaluations the index is built from. So the headline comparison says the flagship remains the better model, and the headline comparison is the least useful thing you can take from these two rows.

What actually separates them is shape rather than rank. D1 Ultra is a model for hard, single-shot judgements: a defect on a circuit board, a damage claim checked against a photo. d1.2 is a model for being asked several things about the same state, at twice the speed, with a context window the board lists at 66k against D1 Ultra's 33k, and with five answer types D1 Ultra does not have. The $0.022 difference is real; whether it is the difference that decides your migration is a question about your request pattern, and that is answerable from the published rows.

The price gap, in the units that matter

Per token, doubling sounds trivial. Per task, it is not the same ratio, because the two models do not consume the same number of tokens to answer the same question. Artificial Analysis models cost per million tasks at current list prices, and the board's own figure is $63 for d1.2 against $133 for D1 Ultra — a 2.1× gap that survives the token accounting.

• Price per million input tokens — $0.02 for d1.2 vs $0.042 for D1 Ultra.

• Modelled cost per million tasks — $63 vs $133, both at current list prices, both from the board rather than from Liquid.

• Decision Index v2.1 — 68.69 (#9 of 25) vs 70.58 (#6 of 25).

• Time per task — 150 ms vs 301 ms. The new cheap model is twice as fast.

• Context window — 66k tokens vs 33k, as listed by Artificial Analysis; Liquid's own launch posts state no window for either model.

• Input modality — text and image for both, per the board. d1.2 adds audio, per the vendor, which the board's spec sheet has not caught up with.

• Brier score, lower being better — 0.164 vs 0.176. The cheaper model is the better-calibrated one.

Read the last line twice. Calibration is the property that makes a probability threshold usable, and the model at half the price wins it. The points the flagship still holds are points about knowing things, not about knowing how sure it is.

Where D1 Ultra still earns its price

A generated two-column scoreboard titled ‘d1.2 vs D1 Ultra - the scoreboard’. The d1.2 column lists Decision Index 69, AA-Paperclip 62.6, When2Call 96.1, AA-Omniscience 60.9, context window 66k and time per task 150 ms. The D1 Ultra column lists Decision Index 71, AA-Paperclip 64.7, When2Call 93.8, AA-Omniscience 64.4, context window 33k and time per task 301 ms. The footer reads ‘Both rows per Artificial Analysis Decision Index v2.1; context windows as listed by the board, not stated by the vendor.’ The OrcaRouter mark is composited bottom right.

Five of the six components favour the older model, and the pattern in them is consistent enough to plan around.

• AA-Paperclip, document decisions — 64.69 for D1 Ultra against 62.60 for d1.2, with 51.3% of documents fully correct against 44.3%. Full-document all-pass is the harsher measure: a fifteen-field invoice with fourteen right counts as a miss.

• EvalSafe-O*NET-AA, workplace judgement — 88.15 against 86.55, and 82.9% accuracy against 79.7%. A 1.6-point difference on the component, which is close to the noise floor of a rerun.

• AA-Omniscience-Decision v1.1, knowledge — 64.39 against 60.94, with 28.0% accuracy against 30.0% and an 88.6% non-hallucination rate against 66.6%. The accuracy row and the abstention row disagree, and the abstention row is the one that dominates the score: D1 Ultra declines far more often and is rewarded for it.

• HLE-MultiChoice v1.1, hard questions — 47.39 against 44.69. On raw accuracy the gap widens the other way, 6.09% against 5.30%, and it is noise on a benchmark where the whole field is under ten percent.

• AA-Symmetry, invariance under rewording — 78.66 against 76.52 on the component score, though d1.2 passes more whole sets, 50.0% against 45.8%. The flagship wins on partial credit and loses on the stricter measure, so this row is not a clean win either way.

If your workload is "one document, one hard question, the answer goes to a human who will act on it," those rows describe exactly the case where the old flagship is still the right call, and the price difference at that volume is a rounding error.

Where d1.2 is not close

One component moves the other way, and it moves a long way.

• When2Call, tool and function decisions — 96.09 for d1.2 against 93.83 for D1 Ultra, with 95.5% accuracy against 93.0%. This is the component that decides whether an agent calls the right function with the right arguments, and it is the strongest single argument for the new model in an agentic stack.

Then there is everything the index does not measure. Five new answer types — a value with the characters it was read from, spans with positions, boxes, counts with ranges, multi-select with per-option probabilities — exist only on d1.2, and the index scores none of them. Their accuracy figures are Liquid's own, on internal test sets built from held-out splits of public datasets, and they are labelled that way everywhere they appear below.

On the vendor's numbers the two new models are close to each other and there is no D1 Ultra row at all, because the task types did not exist when it shipped: 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, and 87% against 90% on picking every object present. The relevant fact for this comparison is not the percentages — it is that the older model cannot express these answers, so a pipeline that needs them is not choosing between two models. It is choosing between one model and a workaround.

The billing rule, which is worth more than the two index points

A headless-browser screenshot of the Artificial Analysis d1-ultra model page, captured 18 November 2026 in English, dated October 2026. The summary panel reads that d1-ultra is among the leading models on the Decision Index and reasonably priced, and notably slow: a Decision Index score of 71 against a median of 66, 301 ms per task against a median of 206 ms, $0.042 per 1M input tokens, and about $133 per 1M tasks against a median of $137. The rank badge reads #6/25 on the Decision Index and #11/25 on cost, and the technical specification lists text and image input with a 33k-token context window.

D1 Ultra bills every question as its own prompt, including every attached image. Ask ten questions about one photograph and the photograph is charged ten times: at Liquid's stated 1.5 tokens per 32×32-pixel patch, a 1024×1024 image is 1,536 tokens, so ten questions carry 15,360 image tokens before a word of text. d1.2 bills a request's text and pictures once, and each additional question adds only its own text.

Work it on a concrete request: a 2,000-token document, one 1024×1024 image, ten questions of about 30 tokens. Under the old rule the request carries roughly 35,700 tokens; under the new one, roughly 3,840. At each model's own price that is about $0.0015 on D1 Ultra against about $0.00008 on d1.2 — the price gap and the billing gap compound rather than cancel. At 250,000 requests a month the difference is roughly $375 against $20.

This is arithmetic on Liquid's own billing description, not a measured invoice, and the ratio collapses to 1× if you ask one question per state. That single question is the whole decision: the more questions you ask about the same document, the less the comparison is about index points at all.

What changes in the client

The endpoint is the same — api.liquid.ai/decisions/v1/systemone — and the model string is the migration. The three original question types keep their shape; responses are still structured JSON with a probability per option and no output tokens, so nothing about your parsing changes. What you gain is additive and does not break anything that already works.

The catch is operational rather than technical: D1 Ultra is not deprecated, both models bill on input tokens only, and Liquid sells both from the same console. There is no forcing function to move, which means the migration will not happen on its own and should be scheduled against a workload you can measure rather than against a launch date.

Running both without committing to either

A word on access before the operational part: d1.2 and D1 Ultra are both served from Liquid's own API, and neither is on our catalogue — this is a comparison, not an availability claim. The reason our catalogue is relevant at all is Jev 1.14, which holds EvalSafe-O*NET-AA at 91.14 where d1.2 manages 86.55, takes the same systemone request shape, and bills $0.042 per million input tokens. If the migration decision comes down to one component you cannot afford to lose, that is the model to test it against.

The split that makes sense follows the corrected component picture rather than the headline. Send multi-question states, anything with an image in it and anything an agent routes on to d1.2 — that is where its context window, its speed and its When2Call lead all point the same way. Keep single hard document judgements and knowledge questions with a definite answer on D1 Ultra until your own sample disagrees. Two routes, one condition each, written by question type rather than hard-coded into three services: that is what the routing DSL is for, and because provider list prices pass through at 0% markup, a change to either Liquid price is visible to the routing the day it happens rather than at the next renewal. Automatic failover then covers the case where one of the two models has a bad afternoon mid-migration.

A headless-browser screenshot of the OrcaRouter model page for typesafe/jev-1.14, captured 18 November 2026 in English, showing the /v1/decisions/v1/systemone endpoint family, Zero Data Retention availability, a $0.04 list price per million input tokens, and measured latencies of 182 ms and 372 ms.

A shadow run costs almost nothing at this scale: a few thousand of your real states through both models, compare probabilities rather than labels, and count the cases where D1 Ultra says 0.85 and is wrong against the same cases on d1.2. The board says the new model is better calibrated overall; your sample is the only thing that says whether it is better calibrated on your data, and that is the number the migration should turn on.