loading
loading
Liquid AI's d1.1 Cuts Decision Costs to $0.02 per Million Tokens Liquid AI's d1.1 narrows the gap to d1-ultra while cutting cost by more than half, keeping input-only billing at $0.02 per million tokens. - Liquid AI released d1.1, upgrading its decision model line between d1 and d1-ultra. - Priced at $0.02 per 1M input tokens, matching d1 and less than half d1-ultra's $0.042. - Claims lower latency than d1-ultra while approaching its decision quality. - Keeps zero output tokens, returning calibrated probabilities for Noul, Choice, and Score questions. - Available now on Liquid Console, with OpenRouter and Vercel integrations coming soon. - Best for routing, moderation, reranking, and LLM-as-judge replacement workloads. Liquid AI has released d1.1, a decision model for classification, routing, scoring, moderation, and automated evaluation. The model is available through the Liquid Console, with OpenRouter and Vercel AI Gateway support planned for a later date. d1.1 costs $0.02 per million input tokens, matching the original d1 and coming in below d1-ultra’s $0.042 rate. Liquid says the new model improves decision quality over d1 while delivering lower latency than d1-ultra, giving developers another option between the base model and the multimodal flagship. Typed decisions without decoding Liquid’s d1 family handles calls that need a label, probability, option, or numeric score. These workloads often use a general-purpose language model to generate a short JSON response, which then requires decoding, validation, and error handling. A decision model returns the structured result directly and reports zero generated output tokens. Each request contains contextual state and one or more typed questions. The API supports three answer shapes: - Noul: Answers a yes-or-no question with a probability from 0 to 1. - Choice: Selects from a defined set and returns a probability for each option. - Score: Rates an item on an ordered scale and returns a weighted position. Direct probability output removes generated JSON from the request path, reducing latency and eliminating malformed-output failures. The resulting probabilities can be thresholded, calibrated, logged, or passed to rules and automation systems. Model errors remain possible, so production deployments still need workload-specific evaluation and monitoring. A cheaper point on the curve At 100 million input tokens, listed model charges would be $2 for d1.1 and $4.20 for d1-ultra, excluding any gateway or platform fees. That difference becomes material for high-volume filters, routers, and evaluation pipelines that may process several decisions per user request. Liquid reports that d1-ultra leads the Artificial Analysis Decision Index and produces calibrated probabilities. Calibration means predictions assigned a 90% probability should be correct in roughly 90% of comparable cases over time. Liquid positions d1.1 near that quality tier, although developers should use their own labeled data to measure the gap. Workloads suited to one-pass decisions Applications that currently call a generative model for a short verdict are the clearest candidates for testing d1.1. Common examples include: - Routing and prioritizing support tickets - Detecting spam, abuse, or policy violations - Classifying documents by topic or industry - Reranking retrieval candidates - Scoring manufacturing quality-control reports - Running LLM-as-judge evaluations - Adding probability-based approval gates to agent workflows Free-form writing, conversation, and multi-step problem solving still require a generative model. Knowledge-dependent decisions also deserve careful benchmarking because the original d1 showed weaker results on tasks that relied heavily on external facts or deeper reasoning. Migration requires one model change Existing d1 integrations can adopt d1.1 by changing the model identifier while keeping the same state, question, and response structures. Teams should rerun evaluation sets before switching production traffic because matching schemas do not guarantee matching probability distributions or optimal thresholds. A useful rollout compares d1.1 with the current model on the same representative traffic and measures: - Accuracy, precision, and recall for each decision class - Calibration at the thresholds used by downstream systems - False-positive and false-negative costs - Median and tail latency under expected concurrency - Input-token usage and total cost per completed workflow Developers can test d1.1 now in the Liquid Console, where new organizations receive trial credit. Liquid has not announced release dates for the OpenRouter and Vercel integrations. Once available, those gateways should make side-by-side traffic tests with d1-ultra easier to run.
document fetched 2026-10-24 15:22