Gemini Decision 1 and the decision-model wave

TL;DR

A decision model reads a state plus a set of questions with fixed options, and returns a probability for each option in one forward pass. It never generates text. In one window, Google shipped Gemini Decision 1 behind a new Decisions API (up to 1,000 yes/no, choice and rubric questions per request, across text, images, audio, video and PDFs, 150-250 ms end to end, $0.10 per million input tokens, zero output tokens billed, about 5x cheaper than Gemini 3.5 Flash-Lite on Google's benchmark suite). Meta released MetaEncoder v1 on Hugging Face, a multimodal embedding model with a Jev-compatible decide() call that scores options by cosine similarity against a softmax at temperature 0.02. inclusionAI released Sing VL S1, a 4B Qwen3.5-based multimodal decision model that leads its size class on the Public Decision Index 0.3 (69.76 text, 73.94 vision). Hugging Face added a decision-model class to Transformers. Google says its own Gemini Notebook already uses the API as a real-time router, serving everyday queries with up to 40% lower latency and reserving deep reasoning for hard ones. Separately, Meta Engineering described Intelligent Delivery, a system that allocates ads-serving compute to the requests with the highest predicted marginal utility; it is the same "score first, then spend" idea applied to compute instead of models.

flowchart LR
  S["State<br/><small>query, doc, video</small>"] --> D["Decision model<br/><small>one forward pass</small>"]
  Q["Questions<br/><small>up to 1,000 per call</small>"] --> D
  D --> P["Probabilities<br/><small>plus confidence</small>"]
  P -->|confident| R["Route or act<br/><small>cheap tier, tool</small>"]
  P -->|ambiguous| H["Fallback<br/><small>big model or human</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class S,Q input
  class D core
  class P loop
  class R exit
  class H err
  linkStyle 3 stroke:#2f9e44,stroke-width:2px
  linkStyle 4 stroke:#e03131,stroke-width:2px

Key points

• Gemini Decision 1: three question types (noul yes/no, choice with 2 to 255 options, score ordered rubric), up to 256K input tokens, latency nearly flat from 1 to 100 questions, lowest calibration error on Google's suite. Gemini Notebook routes queries to model tier and thinking level with it. (Google)
• Price ladder: Jev 4.2 cents per million input tokens, OpenAI's GPT-6 Luna decision model and Gemini Decision 1 at 10 cents; output free on all three. Jev is text/JSON only, OpenAI adds images, Gemini covers all five modalities. Gemini returns hedged probabilities (0.92 rather than 1.0) on yes/no. (Simon Willison)
• MetaEncoder v1: 6656-dim L2-normalised embeddings for text, image, video and documents, with encode, retrieve and decide. Decisions are cosine scores against option descriptions; calibrate() refits the temperature on labelled data. One 80 GB GPU holds it in bf16. (HF)
• Sing VL S1 (4B): one Transformers forward pass per question; 96.8 on BFCL tool-call exact accuracy in its non-skill setting; beats TypeSafe Jev 1.14 (60.70 text) and a 4B Perplexity Decider on the index. (HF)
• Meta Intelligent Delivery: a Resource Core Model (6B sparse, 20M dense parameters) runs once per ad request and predicts outcome, compute and latency per resource setting; a hierarchy of decision engines then moves compute across regions, models and ranking stages. About 10% incremental ranking performance since 2017; about 20% better at finding the highest-impact requests than the prior production model. (Meta Engineering)

Relation to prior wiki pages

• Confirms and completes a pattern. Drex (11-07: open decision models, Max ties GPT-6 Sol at about $0.06 per 1K decisions) and Flide (11-11: an exact serving engine for decision-model request trees, 9.5x over vLLM) made the open and the serving layers. Today adds the hyperscaler API layer (Google), an embedding-based variant (Meta), and the library layer (Transformers). Four layers in six days is a category forming.
• Extends the 11-04 Google cost-per-task study (one classifier question routed a coding suite between Flash-Lite and Flash at lower cost than the best no-router setup). Gemini Notebook is that idea in production.
• Tension with Route Anything (11-05): a perfect per-input router beats the best single model by 5.4 points, but learned routers gained only 0.3. Calibrated decision models make the cheap-path decision more trustworthy. They do not obviously raise the quality ceiling.
• Open: a calibration comparison across Jev, GPT-6 Luna, Gemini Decision 1, MetaEncoder and Drex on the same routing set. Embedding-based decide() (MetaEncoder) versus a trained option-scoring head (Sing VL, Gemini) is an untested design axis.

Concept page: LLM routing.