# d1 Ultra Decision Model: Specs, Pricing, and License By Wendell Achterberg · 2026-10-18 · Source: https://www.activepieces.com/blog/d1-ultra-decision-model-specs-pricing-and-license --- d1-ultra is [Liquid](https://liquid.ai/blog/d1-ultra) AI's most capable hosted decision model, and today it is available through the Liquid API, with OpenRouter and Vercel AI Gateway also listed in the docs. It accepts text and images. The licence and the weights aren't published, so anything that needs either has to wait. Before a data-protection reviewer signs off, they will ask where the data goes. **Every call leaves your environment** and goes to Liquid AI's servers, because Liquid AI hosts the model. [Liquid AI's documentation](https://docs.liquid.ai/lfm/models/d1) describes d1-ultra as a decision model built on a larger open-weight model with Liquid-specific customizations. The page gives no launch date, so I don't state one. ## Understanding d1-ultra and accessing it ### How to get access: weights, API or both You can reach it through the Liquid API, OpenRouter and Vercel AI Gateway, as far as the maker's pages say. That suits a pilot, but a platform that resells you a model has settled your model choice and its price before you open the product, and with a model this new that is the wrong thing to have settled. [Activepieces](https://www.activepieces.com), which MoneyGram, Moneypenny, Alan and FundingSocieties run, works on a provider key you already hold, so d1-ultra's usage is billed at Liquid AI's rate on your provider account and not ours, and the workflow can be reached from Claude, ChatGPT, Cursor or any MCP client. ![Modal dialog for enabling OpenAI as an AI provider in Activepieces Platform Admin, showing API key setup instructions and…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/f98401c7-773d-4cc0-a678-e922f6b6269c/how-webhook-triggers-detect-and-send-real-time-d-5e65f2a0.webp) Because the model sits in one HTTP step, another can replace it without rebuilding the rest. On 2026-10-18 the page stated Liquid API hosted access, text and image inputs, $0.042 per million input tokens and 1.5 tokens per 32x32 image patch, with the license and weights listed as "not published", so you cannot independently verify what the model does. ### License and price as published What you may do with outputs, whether you may redistribute them, and what a procurement file would need to record are all unsettled, because Liquid AI lists the licence and the weights as "not published". Treat the licence as unknown, not permissive. Don't extrapolate a monthly cost from the input price of $0.042 per million tokens, since total spending will depend on how much you actually use the API. It covers only the input side, according to [Console](https://console.liquid.ai/), and there is no separate output price, since d1-ultra is billed on input tokens only and output is free. The image rate of 1.5 tokens per 32x32 patch tells you how to estimate image cost: count patches, multiply, and the total is small per image, so check it against your own volumes during the pilot. ## D1-ultra strengths and performance gaps Almost all of the published evidence comes from the maker or a single benchmark tracker. Treat "state-of-the-art" as a claim to test, not a finding. [Liquid AI](https://liquid.ai/blog/d1-ultra) describes d1-ultra as a multimodal decision model with state-of-the-art performance. It takes text, images or both, plus one or more questions, reads them in one forward pass and returns probabilities without generating tokens. For a workflow, that means the output is a number you can threshold and log, not prose you must parse. ![Image and text going into a box, probability numbers coming out beside a threshold bar and log sheet.](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/8e5107e4-7d93-4fb4-8ff2-38fea018bb1e/d1-ultra-decision-model-specs-pricing-and-licens-afb350ea.webp) ### D1-ultra accuracy limits on BenchLM On [BenchLM](https://benchlm.ai/benchmarks/aa-omniscience-decision-non-hallucination), d1 scores 41% on accuracy and d1-ultra scores 32%,. Expect roughly nine points fewer correct answers from the newer model. I found no independent replication, so your pilot data is the only check you will have. ### Comparison on access, license, price and published evidence Only benchmark figures exist, and every commercial term is blank. Don't put any row into a procurement file until the blanks are filled. | Model | Access route | License | Input price per million tokens | Published evidence | |---|---|---|---|---| | d1-ultra | Liquid API | not published | $0.042 | Maker's blog claim | | d1 | not published | not published | not published | BenchLM 41% accuracy, 1.7% non-hallucination | | d1-3B open-weight | not published | not published | not published | not published | | Jev 1.14 | not published | not published | not published | not published | _Prices and plan limits checked against [liquid.ai](https://liquid.ai/blog/d1-ultra) and [liquid.ai](https://www.liquid.ai/blog/d1-ultra) and [docs.liquid.ai](https://docs.liquid.ai/lfm/models/d1) on October 18, 2026._ Confirm licence and price in writing with Liquid AI before any approval step, since the commercial terms are missing. ## The strongest case for adopting d1-ultra immediately Test it without committing to it: a swappable pilot step lets you find out whether a model built for one job by a credible lab could beat general models on cost, latency and control. ### Why a specialist model could win If you read Liquid AI's pages as a believer, you get four arguments. Each rests on the maker's own statement rather than on anything we've measured, so read them as hypotheses your pilot can confirm or reject. In a routing or approval step, Liquid AI presents d1-ultra as purpose-built for decisions. Liquid AI describes it as reaching its answer in one forward pass with no token generation, which suggests less waiting and less variable output per call than a model that writes out its reasoning. ![Two people at identical counters: one presses a single stamp once, the other writes a long scroll of reasoning that trails…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/cac71e61-1136-4970-908e-0ce0f0f42339/d1-ultra-decision-model-specs-pricing-and-licens-34264a96.webp) Where a workflow sends large volumes of text through a single classification step, a low input price matters most, and Liquid AI states one. It also describes the base model as a larger open-weight one, which hints that self-hosting may become possible. That would keep records inside your own environment, though the licence must be confirmed first. Every one still has to be shown on your own documents. ### Why waiting has its own cost Waiting isn't free. If a decision step is expensive or slow today, every month spent deferring is a month of paying for that. A team that has already built the swappable step has little to lose by running d1-ultra beside its current model on real, logged cases. Meanwhile, the approval chain takes time of its own. Security review, data-handling sign-off and the written licence confirmation can all start now, in parallel with the technical test. A team that begins them only after the results arrive will lose weeks it didn't need to lose. ## The case for swappable workflow steps Starting early is safe only if the new model can be removed again, and a swappable step turns removal into a settings change rather than a project. ### Testing d1-ultra beyond vendor claims Liquid AI's pages tell you only what d1-ultra is meant to do. They don't test it on your documents, your edge cases or your error tolerance. A new model has no months of good behaviour behind it, and the [Code Ninety study](https://www.codeninety.com) shows how pilots usually end: **48.6% stall for six months or more**, 31.4% reach production, and 20% are abandoned, so most new model pilots don't reach production. ![Where enterprise AI pilots end up](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/00f329ac-d71f-41c8-8212-82189ae12239/d1-ultra-decision-model-specs-pricing-and-licens-4a44607c.svg "Source: Code Ninety study") If the model sits in one replaceable step, a failed pilot costs you one step, not a rebuilt process. **Place a named approver ahead of any step that sends, pays or changes a record, and log who approved what.** Until your own results justify removing it, the approver is also your cheapest measure of how often the model is wrong. Set what the maker shows against what your pilot would show, using the table below. Its last column is the evidence that would move this recommendation. | | What the maker shows | What your pilot would show | Evidence that would change the advice | |---|---|---|---| | Vendor benchmark claims | Results on the maker's chosen tests | Results on your own sample of real cases | Your cases fail where the claims say it should pass | | Track record | Recent release, short history | Weeks of logged runs, including failures | Stable output across several weeks of runs | | Cost of being wrong | Not stated for your process | Approver rejection rate and the cost of each miss | Rejections you could not catch before an action took effect | | Data fit | General capability statements | Performance on your formats, language and volumes | Systematic errors on your most common input type | If none of these triggers fires, keep the step swappable and keep the approver in place. ## Connecting d1-ultra in an Activepieces workflow today As of 2026-10-18 Activepieces has nothing specific to d1-ultra, and we haven't run the model, so the pilot goes through a generic step that you can swap out later. Three general routes exist, listed here from simplest to most involved. Before any of them, get approval for sending your data to the endpoint, because a hosted route passes every input to a third party. ### Calling d1-ultra with an HTTP request step Simplest is the HTTP request step, and it's hosted. On the Free plan, which costs $0 and includes unlimited flows, you point Activepieces' standard HTTP step at Liquid AI's hosted API, with the key stored as a connection. ![Flow with HTTP step, key connection and line to a cloud API.](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/2cabe56f-8cfd-4210-8b40-0ab2f17d0b8e/d1-ultra-decision-model-specs-pricing-and-licens-30acac7c.webp) It needs no new software, so removing it later costs one step. Take the request format from Liquid's own documentation, not from our description. ### Using d1-ultra through an AI step or MCP Second comes an AI step with an OpenAI-compatible provider, which can be hosted or self-hosted, depending on what answers the call. It works only if the endpoint speaks the OpenAI request format, and we haven't confirmed that Liquid's does. ![Activepieces workflow builder showing a multi-step automation with the Send HTTP request step selected and its…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/9e36e973-de4a-435f-8912-12e666216b2d/self-host-mistral-ai-enterprise-deployment-guide-3f7fe28d.webp) If it doesn't, an adapter you run yourself would be needed, and that adds a system your auditors will want listed. The third route is an MCP server, also hosted or self-hosted, depending on who runs the server. We found no d1-ultra MCP server, so this means building and maintaining one. It's the heaviest route and not justified for a pilot. ### What the platform costs Cheapest to settle is the platform fee. [Activepieces](https://www.activepieces.com/pricing)' [cloud pricing](https://www.activepieces.com/pricing) is $0 a month on Free, $20 on Plus and $200 on Team, with **Team costing ten times as much as Plus**, so upgrading from Plus to Team costs an additional $180 per month. ![What Activepieces cloud plans cost per month](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/ec561347-9397-44bd-82d2-fde46058b9b4/d1-ultra-decision-model-specs-pricing-and-licens-58c16570.svg "Source: Activepieces") A trial can therefore start without a budget line, and Team is ten times the cost of Plus. Little of the pilot's cost sits in the platform; the model's usage price is separate and not stated here. Next comes the pilot itself: run the same sample through the swappable step and compare the results. ## A Monday plan for piloting d1-ultra safely Run the pilot as six steps in a fixed order, with a named person approving each result before it affects anything real, which means failures cannot propagate without review. 1. Pick one low-risk workflow. 2. Gather 50-100 past cases with known outcomes. 3. Run both models in parallel. 4. Compare accuracy and cost. 5. Keep human approval on every output. 6. Recheck the maker's pages before widening use. Five of the six get working detail in the sections below. The sixth is a standing check: Liquid AI's pages were read on 2026-10-18, and a model this new can change its terms or availability without notice. ### Choose one reversible decision Choose a workflow where a wrong answer can be caught and undone, such as sorting inbound requests into queues or drafting a first-pass classification that a clerk confirms. Avoid anything that commits money, changes a customer record without review, or would need to be explained to an auditor after the fact. Then gather 50-100 past cases whose correct outcome is already known and recorded. Because the answers are settled, nobody has to argue later about what "right" meant. ### Run both models side by side Put d1-ultra in the swappable step beside the model you use now. Send each case to both and store the two outputs, the expected outcome, and the cost of each call in one record. Neither model should see the other's answer, and neither should be allowed to touch live systems during this stage. Keep a person in the approval position throughout. The reviewer sees the output, not which model produced it, and signs off or rejects. That gives you a record of who approved what, which is the first thing an internal audit will ask for. ### Setting d1-ultra pass criteria before testing Write down, before the first run, what counts as a pass: how much accuracy you require, how much extra cost you'll accept for any gain, and which kinds of error are disqualifying regardless of the average. Compare accuracy and cost together. A model that is slightly more accurate but costs noticeably more per case may not justify the change, and the reverse can also hold. If d1-ultra doesn't clear the bar, the swappable step lets you revert in minutes, and nothing else has been rebuilt around it. If it does clear it, recheck Liquid AI's pages, then widen use by one workflow at a time. ## Answering key d1-ultra questions Four questions are left that the pilot cannot settle for you. Each answer below is either what Liquid AI's pages state or is marked as not published. Is Liquid AI d1-ultra open source? Not published: the Liquid AI pages read on 2026-10-18 give no license terms for d1-ultra that I can confirm. Until the license text is in front of you, don't assume commercial use, redistribution or fine-tuning is permitted. Have whoever approves software terms in your organisation read the actual license before the model touches customer data. Can d1-ultra run on my own hardware? Not published: the pages read give no hardware requirements or self-hosting instructions I can confirm. Plan as if hosting were an open question. Keep the model behind a swappable step, so that a later answer about where it runs changes one step and not the whole workflow. Does Activepieces have a d1-ultra integration? Not published: Liquid AI's pages don't list one, and the 741 integrations counted in Activepieces' open source monorepo include none that we know of built for it, so Liquid AI isn't available as a pre-built integration. Until the maker states otherwise, the practical route is a generic HTTP request step pointed at whatever endpoint Liquid AI documents. That's the reason to isolate the model in one replaceable step. ### How do I tell if d1-ultra beats my current model? Run both on the same sample of your own real inputs, and score the outputs blind against criteria you wrote down beforehand. Independent benchmarks are limited for d1-ultra, and the maker's own claims are not evidence about your workload. Price is published at $0.042 per million input tokens with no output charge, so record cost per run during the pilot to check it against your volumes. A result you can reproduce and show to an approver is the only comparison that should decide the question. ## Related reading - [Cloudflare Clef: Open Decision Models Explained](https://www.activepieces.com/blog/cloudflare-clef-decision-models-for-agentic-workflows) - [Reconstructing an AI Decision for a Customer Dispute](https://www.activepieces.com/blog/reconstructing-an-ai-decision-for-a-customer-dispute) - [Xing4.0-29B-A4B: Agentic MoE Model Guide (2026)](https://www.activepieces.com/blog/xing40-29b-a4b-agentic-moe-model-guide-2026) ## References - [Code Ninety study](https://www.codeninety.com) - [Activepieces](https://www.activepieces.com/pricing) - [BenchLM](https://benchlm.ai/benchmarks/aa-omniscience-decision-non-hallucination)