LATEST
AI Tools

Liquid AI d1.2 Is Live: $0.02 Decisions, Audio Input, Top Score

Updated November 17, 2026
12 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of a digital checklist being marked off on a tablet with a stylus pen

Quick Answer & Key Takeaways

  • • Liquid AI released d1.2 and d1.2-ultra on November 16, 2026, decision models that return calibrated probabilities instead of generated text.
  • • d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, the top score among decision models.
  • • Both models cost $0.02 and $0.042 per million input tokens, with no charge for output.

Most AI models answer a question by writing. Liquid AI's decision models answer with a number instead: a probability, a value read from a document, or a box drawn around an object in a photo. On November 16, 2026, the company released d1.2 and d1.2-ultra, the newest versions of that idea, adding audio input and five new tasks. The headline number is a benchmark score: d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, the highest score any decision model has recorded on that index. The price is the other headline. d1.2 costs $0.02 per million input tokens and d1.2-ultra costs $0.042, with output free because neither model generates tokens. What follows is what a decision model is, what the new version can do, what the benchmark does and does not prove, and what a million decisions would actually cost.

What a decision model actually does

A decision model is built for structured calls, not conversation. Instead of generating text token by token, it evaluates a situation and returns calibrated probabilities across a fixed set of outcomes in a single call, with zero generated tokens. That is how Liquid AI's documentation describes the class, and it is the key to understanding why the pricing looks so different from a chat model. There is no output to bill because there is no output text.

The practical uses are the unglamorous work inside software: routing a support ticket to the right team, scoring a bug report by severity, checking whether a message contains personal data, approving or blocking an agent's tool call, and replacing a large language model that was being used as a judge. Liquid lists classification, routing, scoring, triage, content moderation, guardrails, and model routing among the fits. A question such as "Is this message spam?" comes back as a probability, for example 0.92, which lets a developer set a threshold and send low-confidence cases to a human for review.

This is a category with real competition now. Google, Microsoft, OpenAI, and Perplexity all run decision-model products in 2026, and this site has already covered two of them: Drex Max v1, a free 27B decision model, and Google's Gemini Decisions API. Liquid's entry is the newest, and its pitch is accuracy per dollar rather than raw capability.

What is new in d1.2

The release adds five tasks to the Liquid API, all in the same single-call format as the older question types. According to Liquid's launch post, published November 16, 2026, the new tasks are:

  • Read a value from the input. Ask "How much is due?" about an invoice and the model returns the value with the exact characters it read it from. If the invoice does not show the value, it answers "not found" instead of guessing.
  • Find every passage of a kind in a text. Ask for every piece of personal data in a note and the model returns each passage, such as a name, a phone number, and an email address, with its position in the text.
  • Locate things in a picture. Ask where the signature is on a scanned form and the model returns a rectangle around it, or "not found". Ask for every tree in a photo and it returns a rectangle around each tree and the count.
  • Estimate a number, with a range. Ask how many people are in a photo and the model returns its best guess, for example 12, with a range such as 10 to 14.
  • Pick every option that applies. Ask which topics a support message raises and the model returns each topic that applies, such as billing and delivery, each with its own probability.

Audio is the other addition. Both models now take audio as input, which can be the thing asked about, for example which language someone speaks or whether there is a siren in the background, or the question itself, spoken instead of typed. On held-out recordings, Liquid reports that both models answer 95% to 97% of questions about which sound they hear or which language is spoken, across 43 languages, with three to six options per question. Those figures are company-reported, and the close confusions Liquid names, such as Spanish and Portuguese, are the honest kind of detail that makes the number easier to trust.

Two specification changes matter for builders. The d1.2-ultra context window grows from 33,000 to 66,000 tokens, and average task latency falls from 301 milliseconds to 218 milliseconds, a cut of about 28%, according to launch coverage. The pricing, notably, does not change: d1.2 and d1.2-ultra keep the same per-token rates as their predecessors.

The benchmark score, and what it does not prove

The Artificial Analysis Decision Index v2.1 evaluates decision models on six benchmarks, covering decisions and extraction on resumes, receipts, and research papers, plus tool calling and hard knowledge questions. It also scores the probability a model assigns to each answer, so a confident wrong answer costs more than a hesitant one. On that index, d1.2-ultra scores 75, the highest score among decision models and two points ahead of the previous best, according to Liquid's launch post. d1.2 scores 69. Both scores are corroborated by independent launch coverage.

Calibration is where Liquid's case is strongest. A probability is only useful if it is correct: when a model says 80%, it should be right about 80% of the time. Across 13 benchmarks and 22,505 questions answered by all five models Liquid compared, d1.2-ultra's calibration error is 0.8 points, against 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 OpenAI's Decisions API. A well-calibrated model lets a team auto-approve high-confidence answers and route the rest to people, which is the whole point of using a decision model in production.

The new tasks carry a clear caveat, and Liquid states it openly: no public benchmark measures the five new tasks yet, so the company evaluated them on internal test sets built from held-out splits of public datasets. The reported results are:

Taskd1.2-ultrad1.2
Read a value from a receipt98%96%
Find every piece of personal data in a document (precision / recall)94% / 93%92% / 88%
Locate an object described in a photo94%94%
Estimate the number of objects in a photo, within one93%92%
Pick every object present in a photo87%90%

One row deserves attention. On picking every object present in a photo, the cheaper d1.2 beats d1.2-ultra, 90% against 87%. That is a useful reminder that a higher index score does not mean a win on every task, and it is a reason to test both models on your own data before committing to the pricier one. Until a public benchmark covers these tasks, the numbers above are Liquid's own, not independent results.

What it costs

The pricing is simple, and it is confirmed in both the launch post and the documentation. Both models bill input tokens only, because decision models do not generate tokens.

ModelPer million input tokensOutput charge
d1.2$0.02 (about £0.015, C$0.029)Free
d1.2-ultra$0.042 (about £0.032, C$0.060)Free

Converted figures use November 17, 2026 market rates of 1 USD = 0.758 GBP and 1.43 CAD, and are approximate. At that scale, one billion input tokens costs $20 on d1.2, about £15.16 or C$28.60, and $42 on d1.2-ultra, about £31.84 or C$60.06.

Images are billed by patch, not by file. An image costs 1.5 tokens per 32-by-32-pixel patch at the text rate, so a 1024-by-1024 image costs 1,536 tokens. That works out to about $0.000031 per image on d1.2 and about $0.000065 on d1.2-ultra, or roughly $0.031 and $0.065 per thousand images. A request bills its text and pictures once, and each extra question adds only its own text, so asking ten questions about one invoice costs little more than asking one.

A worked example shows how small the bills get. Take a support team routing one million tickets a month, with each ticket and its routing question totalling about 500 input tokens. That is 500 million tokens, which costs $10.00 a month on d1.2 and $21.00 on d1.2-ultra. The arithmetic is ours, using Liquid's published rates; your own token counts will differ, but the order of magnitude is the point. At these rates, the model is rarely the expensive part of a decision pipeline. The integration work is.

How d1.2 compares

Decision-model pricing now spans two very different shapes: per-token billing and per-request billing. The table below uses list prices checked by Parallel on November 6, 2026, alongside Liquid's own rates.

Provider and modelInput priceOutput charge
Liquid d1.2$0.02 per million tokensFree
Liquid d1.2-ultra$0.042 per million tokensFree
Perplexity pplx-decider-v1.2$0.02 per million tokensFree
Jev 1.14$0.042 per million tokensFree
OpenAI Decisions API (GPT-6 Luna)$0.10 per million uncached input tokensFree
Parallel Decisions API$1 per 1,000 requests, up to 100 questions eachNot applicable

On price alone, d1.2 ties Perplexity's decider for the lowest per-token rate in this table, and d1.2-ultra matches Jev's rate while scoring higher on the Decision Index. OpenAI's Decisions API costs five times more per input token than d1.2, though its 7.2-point calibration error is the number that should worry a buyer more than the price. Parallel's per-request pricing is a different shape entirely and suits teams that ask many questions per call, since one request can carry up to 100 questions with live web search included.

The benchmark picture, where scores are published, puts d1.2-ultra first at 75. The models behind it in Liquid's own comparison are Jev 1.14 and Perplexity's Decider v1.3, with GPT-6 Luna further back. Price and score do not move together here: the most accurate model in the comparison is also among the cheapest, which is unusual enough to be worth stating plainly.

How to try it

Both models are available now on the Liquid API. Call the models as d1.2 and d1.2-ultra, with an API key from the Liquid Console under Dashboard, then API Keys. The new answer types use the same endpoint and request format as the existing ones, so teams already calling a d1 model do not need to relearn the API. The full reference is in Liquid's decision-model documentation.

OpenRouter and Vercel support is listed as coming soon, with no date given. That matters if your stack already routes through one of those platforms, since today the only hosted route is Liquid's own API. Liquid also says open-weight versions with equivalent capabilities will be released in the future, again with no date. The open-weight models available now, d1-2B, d1-omni-2B, and d1-omni-600M, are the previous generation, free to run on your own hardware under the LFM Open License, with commercial use free below $10 million in annual revenue. For the open-weight side of the same lab, see our review of Liquid AI's LFM3 models.

The catches

Three things deserve a pause before building on d1.2. First, the predecessors are gone: d1.1, d1-ultra, and the original d1 are deprecated, according to the documentation. Anyone still calling them should migrate to d1.2 or d1.2-ultra rather than wait for a forced cutoff.

Second, the open-weights story is a promise, not a download. The hosted models are the only way to use d1.2 today, and Liquid has given no date for open-weight versions with the same capabilities. Teams that need to self-host for privacy or compliance should look at the existing open-weight d1 models, or at Drex Max v1, which ships as a free 27B model, rather than plan around a release that has no date.

Third, the new-task numbers are internal. Liquid is transparent about this: no public benchmark covers the five new tasks, and the accuracies come from its own test sets. The Decision Index score is independently run by Artificial Analysis, but the receipt-reading, PII-finding, and object-location figures are not. Test on your own documents and images, and calibrate thresholds on your own data, which Liquid's own documentation recommends.

Who should use it

d1.2 fits teams that already know the question they want answered. A support operation routing tickets, a finance team pulling values from invoices, a moderation queue sorting flagged content, or an agent framework that needs a cheap approval check before a tool runs will get the most from it. The audio input adds call routing and spoken-intent detection to that list.

It is the wrong tool for anything that needs prose. Decision models do not write, summarise, chat, reason through multi-step problems, or generate code. If the answer is a paragraph, use a language model. If the answer is a choice, a score, a yes or no, or a value copied from a document, a decision model does the job for a fraction of a cent, and d1.2-ultra currently does it with the best-published score and calibration in its class. For most teams, d1.2 is the right starting point, with the ultra model reserved for the calls where accuracy matters most.

Frequently asked questions

What is a decision model?

A decision model answers structured questions with probabilities instead of generated text. You send it a situation and a set of possible answers, and it returns how likely each answer is, in a single call with zero generated tokens. It suits classification, routing, scoring, and extraction, not free-form writing.

How much does Liquid d1.2 cost?

d1.2 costs $0.02 per million input tokens and d1.2-ultra costs $0.042, with no charge for output. At November 17, 2026 rates, that is about £0.015 and C$0.029 per million tokens for d1.2, and about £0.032 and C$0.060 for d1.2-ultra. Images are billed at 1.5 tokens per 32-by-32-pixel patch.

Is there a free version of d1.2?

There is no free hosted tier published in the launch post or the documentation, though Liquid points to its pricing page for current prices and available credits. The free route is self-hosting: the open-weight d1-2B, d1-omni-2B, and d1-omni-600M models run on your own hardware under the LFM Open License. Open-weight versions of d1.2 itself are promised with no date.

How is d1.2 different from d1.2-ultra?

d1.2-ultra is the more capable model: it scores 75 on the Artificial Analysis Decision Index v2.1 against 69 for d1.2, and its calibration error is 0.8 points against 1.2. It costs just over twice as much per token. d1.2 is the faster, cheaper default, and it wins one of the five new tasks outright, picking every object in a photo at 90% against 87%.

Can a decision model replace ChatGPT?

No. A decision model cannot write text, hold a conversation, summarise a document, or write code. It answers one structured question per call. The sensible setup is to use a decision model for the routing, scoring, and extraction steps, and keep a chat model for anything that needs prose.

Where do I get an API key?

Create one in the Liquid Console at console.liquid.ai, under Dashboard and then API Keys. Both models are available now on the Liquid API, called as d1.2 and d1.2-ultra. OpenRouter and Vercel support is listed as coming soon.

Liquid AI's d1.2 is a narrow tool that does its narrow job well: it answers structured questions cheaply, with the best-published benchmark score and calibration in its class, and it now reads documents, pictures, and audio. The new-task numbers need independent testing before they can be taken at face value, and self-hosters will have to wait for open weights. For teams already paying a chat model to make yes-or-no calls, switching those calls to d1.2 is the cheapest experiment in AI infrastructure this month, and the documentation to start is linked above.

Sources

12 min read
65 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more