Liquid AI has released d1.2 and d1.2-ultra, decision models that skip token generation entirely: given a state plus typed questions, they return calibrated probabilities for every option in a single call, adding five new answer types including value reading, span finding, box drawing, range estimation and multi-select. d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, two points ahead of rivals at 73, with a 0.8-point calibration error versus 3.7 for Perplexity Decider v1.3. The trade-off: d1.2 scores just 69, trailing eight models, and parameters, latency and weights remain undisclosed.
Calibration Error
校准误差
Metric measuring the gap between predicted confidence and actual accuracy; lower values indicate more reliable probability estim
Curated Intelligence
Google has launched the Gemini Decisions API, powered by Gemini Decision 1, its first multimodal decision model built for closed-ended choice tasks. It natively supports Yes/No, multiple-choice with up to 255 options, and rubric-based scoring, handling 1,000 questions per call at $0.10 per million input tokens with output tokens free of charge. The model leads benchmarks with the lowest calibration error, marking a shift from generative output toward structured decision interfaces.
Google launched Gemini Decision 1 and the accompanying Decisions API, which skip token-by-token generation and instead score a predefined set of candidate answers in a single forward pass, returning the decision, probability and confidence. The model handles yes/no judgments, multiple choice with up to 255 options and rubric-based scoring, processes up to 1,000 questions per request with 150-250 ms end-to-end latency, and is priced at $0.10 per million input tokens with free output. Google claims first place on 7 of 9 benchmarks with up to 92.2% accuracy, and says internal teams already use it in Search and product features.
Google released the Gemini Decisions API, powered by Gemini Decision 1, shifting the LLM from text generation to structured scoring: a single forward pass scores up to 1,000 yes/no, multiple-choice or rubric questions with calibrated probabilities and confidence values, always emitting zero output tokens. It leads 7 of 9 accuracy tests on Google's own benchmark table, hitting 92.2% on JevBench with a calibration error of 0.0250, roughly half that of TypeSafe Jev 1.14, though it trails on When2Call at 79.9% versus 82.0%. Pricing is $0.10 per 1M input tokens with free output and 150-250 ms end-to-end latency.