Liquid d1: Decision API, Pricing and Jev Comparison

Compare Liquid d1 with Perplexity’s current v1.2 or Flash API on the same labeled routing workload. Perplexity can share state processing within a question group; Liquid bills each question’s prompt separately. Use returned input usage and retries to compare cost per accepted decision. Perplexity’s historical Apache 2.0 v1.1 checkpoint has a separate self-hosting budget.

Liquid introduced d1-ultra on October 15 for text and image decisions, including named noul, choice and score questions. Use model d1-ultra at https://api.liquid.ai/decisions/v1/systemone through Liquid’s native API. Start with the official request example and your own labels before switching an existing d1 workload.

d1-ultra costs $0.042 per million input tokens with free output. Each question is billed for its own prompt, including the text and all supplied images. Three questions over the same content therefore need three input budgets; Perplexity’s shared-state grouping rules do not transfer to this endpoint.

Liquid counts 1.5 tokens per 32×32 image patch, so a 1024×1024 image contributes 1,536 tokens per question. For an assumed 100,000 questions with 2,000 billed input tokens each, the token bill is $8.40. Use measured request usage for real workloads and include retries.

The launch announcement describes a larger open-model base and reports vendor evaluation results. That is not a release of d1-ultra weights. The smaller d1 checkpoints below remain different deployment options. Liquid’s announcement places Vercel and OpenRouter access in a coming-soon section; use the native API path unless your gateway currently lists the exact ultra model.

Choose the output shape your workflow needs. A support router needs named categories; an urgency sorter needs ordered levels; a yes/no check needs a probability. In each case, write criteria that distinguish neighboring answers. “Billing handles charges and invoices” gives a router more useful information than a bare department name.

In the illustrated checkout example, the hypothetical Choice probabilities are technical 86%, billing 10% and triage 4%. Urgency levels are indexed 0 (can wait), 1 (today) and 2 (time-sensitive); probabilities [0.05, 0.15, 0.80] give a weighted Score of 1.75. Noul is 0.94 for “Is this a bug report?”.

A Noul value near 0.5 expresses uncertainty between yes and no, rather than medium severity. Use Score for severity, with descriptions of each level. For routing, examine the Choice distribution when two departments both fit. Decide how your application handles that ambiguity before connecting the answer to an action.

Create a key in the Liquid console and set LIQUID_API_KEY in your terminal environment. The request below asks which queue should receive a duplicated-invoice ticket. It follows Liquid’s current documented endpoint, /decisions/v1/systemone, and uses d1:free. The result contains a named answer and probabilities; your application chooses what to do next.

Start with a small set of tickets with known destinations. Include short messages, contradictory descriptions and requests spanning two teams. Compare the predicted category and the runner-up probabilities with those labels. Keep uncertain cases in a triage queue while you adjust the criteria.

Send a local JPEG as a Base64 data URL in images alongside state and questions. Use model d1 on Liquid’s direct HTTP endpoint: d1:free rejects images, and the TypeSafe SDK currently has no image input support. The example asks two independent questions about the same board; your application reads answers.defect.noul and answers.board_type.choice.

Liquid’s official Playground demonstrates inspection on four VisA production-line categories: circuit boards, candles, cashews and chewing gum. Liquid reports 85–97% accuracy in those examples. It also demonstrates reading a Wordle board from a screenshot. These publisher examples show concrete visual tasks; choose questions and labeled images that match the parts or screens in your workflow.

All questions receive all images in the request, in the supplied order. For a reference-and-inspection pair, identify image 1 and image 2 in the state and ask about the inspected part. JPEG, PNG, WebP and GIF are supported, with at most eight images, a JSON body below 4.5 MB and at most 10,000 patches of 32 × 32 pixels across all images. Each image’s longer side must be no more than 100 times its shorter side. Compress or resize before Base64 encoding when the body is too large.

For visual decisions, call Liquid directly with paid d1. For text-only routing, the free SKU or either gateway is an alternative. Send the images array through Liquid’s direct endpoint.

Liquid’s current pricing page and OpenRouter’s liquid/d1 listing both charge $0.02 per million input tokens and cached input tokens, with free output, checked October 18, 2026. Cached input has no lower rate for this SKU. OpenRouter lists a roughly 66K-token context. Earlier October 2 and October 5 quotes were $0.04 per million input tokens; use the current listed rate for a new budget. The separate text-only d1:free offer is a different SKU.

At the OpenRouter input rate, 100,000 calls averaging 1,000 billed input tokens cost $2. If the average grows to 5,000 tokens, the same call volume costs $10. Include the questions and criteria in the input count as well as the state. Adding a generation step to explain or act on a decision adds that model’s cost to the workflow.

Liquid counts 1.5 input tokens per 32 × 32-pixel patch. For each image, round width ÷ 32 and height ÷ 32 up to whole numbers, multiply them, then multiply by 1.5 and round up again. A 1024 × 1024 image has 1,024 patches, or 1,536 billed tokens per question. Two questions about that same image count 3,072 image tokens: 3,072 ÷ 1,000,000 × $0.02 = $0.00006144 for the image portion of one request. At 100,000 such requests, the image portion is $6.144, plus the billed text for each question. This example uses Liquid’s paid d1 rate checked October 18, 2026; the total follows the image dimensions and question count above.

Combining questions reduces network round trips, but the supplied image is billed for every question. Read usage.input_tokens to include the state, instructions and criteria actually billed. A separate text-generation step adds its own input and output fees.

Liquid’s 19–200× lower-cost comparison covers six applications run once per model on October 5, at list prices without prompt-cache discounts and with the comparison models at default reasoning settings. It is evidence about those applications, rather than a saving to apply to every visual workload.

LocalLLaMA’s typed-decisions benchmark measured d1 through Liquid’s API on September 30: 400 cases, five questions per request, 2,000 decisions and zero errors. Its hosted rows report sequential client-to-answer median latency. The same client code was used for the Jev row, measured on September 18.

The test labels are averages of three samples from a roughly 4B-class teacher. Accuracy here measures agreement with that teacher. d1’s 0.742 score exceeds Jev’s 0.727 in this test, while meraGPT Decider 1 scores 0.768. This gives you a reproducible comparison to investigate, rather than an accuracy forecast for your ticket queue.

The per-type results also matter: d1 scores 0.840 on Noul, 0.732 on Choice and 0.677 on Score. A workflow built around rubric scoring deserves a separate trial from a yes/no filter. Evaluate errors by category, since one average can hide the decisions that are expensive to get wrong.

Liquid d1 and Jev both offer hosted typed decisions, so a comparison can reuse your task descriptions and labeled cases. Measure response time from the application, track routing errors and count total billed inputs. Keep the provider and model ID in the results so a free-tier trial does not become an assumed permanent operating price.

Jev 1.13 and 1.14 also accept experimental image requests with TypeSafe’s image header; Box questions require 1.14. Liquid’s paid direct d1 route has its own image format and billing. Compare the same visual task and channel, or use Clef’s visual-input and self-hosting options.

For local inference, Laya and CLM-8B offer downloadable checkpoints with different serving paths. CLM scores candidate actions and can reuse candidate embeddings. Laya exposes decision outputs through its own runtime. Local deployment adds device memory, serving and idle-time costs; use a measurement from the exact model and workload in the decision calculator.

Liquid now publishes LiquidAI/d1-3B and the experimental LiquidAI/d1-omni-600M as downloadable decision models. These checkpoints are separate from the hosted d1 and d1:free API SKUs. d1-3B accepts text and images; the experimental omni checkpoint also adds audio inputs. Use each checkpoint’s model card for its supported request format and runtime.

The weights use the LFM Open License v1.0, rather than Apache 2.0. Its commercial-use condition is tied to annual revenue: the license defines the threshold as $10 million or more, and Section 5 excludes commercial use by a Legal Entity exceeding that threshold. A Legal Entity includes entities under common control, not just one deployment team. Below the threshold, commercial rights are granted subject to the license; above it, obtain separate permission from Liquid before commercial use. The revenue threshold does not apply to a qualified nonprofit’s noncommercial or research use. They return typed decisions in a single forward pass with zero generated output tokens. Budget device memory, serving time, idle capacity and input processing; local costs follow the compute used for each decision.

For these local checkpoints, start with a small labeled workload, include refusal handling and measure the runtime configuration. Compare accepted decisions and complete serving cost with the existing hosted service.