Liquid AI Says Its $0.042 Model Tops OpenAI on Calibration

Liquid AI released two decision models, d1.2 and d1.2-ultra, on 16 November, and says d1.2-ultra’s calibration error is 0.8 points against 7.2 for GPT-6 Luna, the only model on OpenAI’s Decisions API. That is a gap of 9 times, by our arithmetic from the figures in Liquid’s own post.

Liquid prices d1.2-ultra at $0.042 per million input tokens and d1.2 at $0.02, the same ultra price as its October model. At those rates, reading a million 1024 by 1024 images costs about $64.51 on d1.2-ultra and $30.72 on d1.2, image tokens only, by our arithmetic.

Business Pill · Verification Cost

A short explainer on verification cost, what it takes to check that an answer is right. It teaches the general idea and does not discuss Liquid AI or OpenAI.

The key insight: By our arithmetic, Liquid AI’s own figures put GPT-6 Luna’s calibration error at 9 times d1.2-ultra’s, at a price of $0.042 per million input tokens, or about $64.51 per million 1024-pixel images.

What Liquid Released

Liquid’s post says the two models add audio input and five new tasks: reading a value with “the characters it read it from”, finding every passage of a kind, drawing boxes around objects, estimating a number with a range, and picking every option that applies.

If an invoice does not show a value, the post says, the model answers “not found” rather than guess. Both models bill input tokens only, and Liquid says a 1024 by 1024 image costs 1,536 tokens at the text rate.

Liquid says d1.2-ultra “scores 75, the highest on the index and two points ahead of the previous best decision model” on the Artificial Analysis Decision Index v2.1. The post does not give d1.2’s index score.

Calibration error of five decision models as Liquid AI reports it in its 16 November 2026 post (13 benchmarks,
Calibration error of five decision models as Liquid AI reports it in its 16 November 2026 post (13 benchmarks, 22,505 questions answered by all five). Liquid’s own measurements, not independently verified; the 9x ratio is our arithmetic.

The Calibration Claim

A decision model returns a probability with each answer. Liquid’s post explains: “When it says 80%, it should be right about 80% of the time, which lets you set a threshold.”

Across 13 benchmarks and 22,505 questions, Liquid reports calibration errors of 0.8 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. By our arithmetic, Perplexity’s figure is 4.6 times d1.2-ultra’s and Jev’s 3.4 times.

These are Liquid’s own measurements; OpenAI’s documentation describes its Decisions API as returning typed answers “about 10x faster than the Responses API” and lists gpt-6-luna as “the only currently supported model”.

Grouped bar chart: Liquid internal test accuracy, d1.2-ultra vs d1.2: read receipt value 98 vs 96, locate object 94 vs 94, count within one 93 vs 92, pick all objects 87 vs 90
Accuracy of d1.2-ultra and d1.2 on the five new task types, on internal test sets built from held-out splits of public datasets, as Liquid AI reports it on 16 November 2026. Liquid’s own measurements, not independently verified.

What the Price Buys

d1.2-ultra costs 2.1 times d1.2 per token, by our arithmetic. On Liquid’s internal test sets, ultra reads receipt values correctly 98% of the time against 96% for d1.2, while d1.2 scores higher at picking every object present in a photo, 90% against 87%.

In October, Liquid said d1-ultra led other decision models on the index “by at least 8 points”. This month it gives a two-point lead; the October post does not name an index version, so the two leads may not be on the same scale.

How the low end of the AI market splits between tokens and spend is in our OpenAI low-end tokens piece, and another start-up pricing per decision is in our RightNow piece.

The Structural Read

As we read Liquid’s post, the release competes on verifiable output: each answer comes with a probability and, for values read from a document, the characters it came from, so a buyer can route low-confidence cases to people.

As we read the two Liquid posts, the price of the ultra tier did not change between October and November while audio input and five task types were added; the lead it claims on the index went from at least 8 points to two.

Liquid AI, d1.2 launch post, 16 November 2026

“For example, you can automatically send tasks where the model has low confidence to human review.”

Three Implications

FOR THE LOW END OF THE MARKET Token share against spend share is in our OpenAI low-end tokens piece.

FOR PER-DECISION PRICING Another start-up pricing per decision is in our RightNow piece.

RELATED Liquid’s d1.2 post is the primary source for every model figure in this piece.

The Business Engineer Lens

This story maps onto the Business Engineer framework AI and the Great SaaS Bifurcation.

The framework argues: “What emerges is a barbelled market structure where software companies must choose between racing to the commoditization floor or climbing to the embedding ceiling.”

As we read it, Liquid prices at the floor, fractions of a cent per image, and argues on calibration rather than general capability. The framework predates the release.

What Is Not Established

Every benchmark and calibration figure here comes from Liquid’s post; Artificial Analysis’s own page and OpenAI were not checked for these numbers, and we did not test the models ourselves. The new-task accuracy comes from what Liquid calls internal test sets.

An X post about the release gave d1.2 an index score of 69 and a four-point jump; Liquid’s post states neither, so we do not use them. Liquid and OpenAI were not contacted for this piece.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

AI and the Great SaaS Bifurcation

The Business Engineer analysis of why software splits between a commoditization floor and an embedding ceiling.

Read the premium analysis →Get 50% off →

The Bottom Line

Liquid AI says d1.2-ultra tops the Decision Index at 75 and has a calibration error of 0.8 against 7.2 for OpenAI’s GPT-6 Luna, at $0.042 per million input tokens, or about $64.51 per million 1024-pixel images by our arithmetic.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Model features, prices, index scores, calibration errors and task accuracy are from Liquid AI’s posts of 16 November and 15 October 2026 and are Liquid’s own claims; the description of OpenAI’s Decisions API is from OpenAI’s documentation, read on 17 November. Ratios and per-image costs are our arithmetic. Liquid AI and OpenAI were not contacted for this piece.

Sources: Liquid AI, d1.2 launch post, 16 Nov 2026 · Liquid AI, d1-ultra launch post, 15 Oct 2026 · OpenAI API docs, Decisions guide

Scroll to Top
FourWeekMBA