Gemini Decision 1

Google DeepMind's hosted decision model, announced Nov 11 — the third hyperscaler into the class after OpenAI and Microsoft, and the most multimodal hosted entrant: text, images, audio, video, and PDFs natively in one input (256K context), up to 1,000 questions per request mixing noul / choice (2–255) / score, scored in one forward pass with per-answer probabilities and confidence — 50–100 ms of model compute, 150–250 ms end-to-end, near-flat from 1 to 100 questions. $0.10/M input tokens, output billed at zero, and the shared input billed once per request no matter how many questions ride it. Google claims the lead across its own routing/reasoning/ranking/multimodal suite at the lowest calibration error and ~5× cheaper than Gemini 3.5 Flash-Lite; its first independent row — #9 of 19 on Artificial Analysis's v2.1 board, behind Jev 1.14 — is solid, not the claim. Already the control plane inside Google: Notebook's model-tier router (up to 40% lower latency), Stitch's rapid canvas, CC's 200-option emoji picker, and Search's MCP-tool safety screen; adapters rolling out across OpenRouter, LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, with Vertex "coming soon." Note the vocabulary: Google shipped noul / choice / score — the TypeSafe dialect is becoming the standard wire.