Treg has launched an open-source jobs API that lets AI agents search employer job boards, score fit against a resume and locate the hiring manager for a role. Fit scoring is handled by TypeSafe AI's decision model Jev at under $0.001 per evaluation, bringing a full job hunt to roughly $0.03 with no markup.
Decision Model
决策模型
Small models for structured question answering that output typed answers with calibrated probabilities instead of free-form text
Curated Intelligence
PostHog's engineering team shares production lessons from using decision models such as TypeSafe's Jev, reporting they are up to 193.6x faster and 444.6x cheaper than frontier LLMs on structured question-answering tasks. Unlike free-form LLM prose, decision models return typed answers with calibrated probabilities, making uncertainty explicit and enabling them to gate expensive LLM calls.
More than one in four requests on Vercel's AI model router went to decision models in October, a category invented only a month earlier. These models skip output tokens and instead make fast, cheap choices among preset options with confidence scores, suiting latency-sensitive production workloads. TypeSafe hit a $7.5 billion valuation on its Jev family, but low barriers have spawned dozens of rivals, pushing customization to the fore.
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.
Google beat Anthropic to market with Gemini Decision 1, a calibrated multimodal decision model unveiled at the AIE CODE event, supporting up to 1,000 questions per call at $0.10 per million input tokens with roughly 200ms end-to-end latency. The launch marks decision models as a distinct production-ready product category and gives Google an early lead in the race with Anthropic.
Nace AI has launched Drex, a document parser that uses decision models to compare each OCR block against the original page image and sends only mismatches back for correction. It scores 0.834 on the Parse Index, beating GPT-6 Astra's 0.813, at $7 per 1,000 pages versus $70.57.
Google has entered the decision-model category with Gemini Decision 1, joining Jev and OpenAI. The API supports yes/no, choice, and rubric-scoring questions, and is the only one of the three to handle text, images, audio, video, and PDFs. Pricing matches GPT-6 Luna at 10 cents per million input tokens with free output, still above Jev's 4.2 cents.
Google DeepMind has launched the Gemini Decisions API and its underlying Gemini Decision 1 model, purpose-built for closed-ended decisions such as yes/no, multiple choice, and scale ratings. It handles up to 1,000 independent questions in a single pass across text, images, audio, video, and PDFs, returning probabilities with confidence values in 150-250 ms. Google says it already uses the API internally for dynamic model routing in Gemini Notebook, design prototyping in Stitch, and safety checks in Search, signaling that decision models are becoming a mainstream product category.
Amazon's internal ticketing system swapped Claude Sonnet 4.6 for Strands Decider, a decision model launched a month ago, achieving comparable labeling quality at 98% lower token cost and saving millions of dollars annually. The key insight is reframing open-ended text generation as a constrained decision problem over fixed labels, letting a specialized decision model replace a general-purpose LLM.
Hugging Face has introduced Decision Models, which move classification and ranking tasks away from LLM text generation into a typed, non-generative scoring interface. The approach supports multiple question primitives, runtime-defined categories, and specialized training for more reliable, calibrated decisions in scenarios such as support ticket routing.