Gemini Decision 1 Returns Probabilities, Not Text, at $0.10/M

Google launched Gemini Decision 1 (gemini-decision-1) on November 11, 2026 through a new Gemini Decisions API. Google prices it at $0.10 per million input tokens and bills zero output tokens. The model never decodes text; it scores up to 1,000 typed questions against a single shared input in one forward pass and returns calibrated probabilities. The target is the small branching calls inside agent pipelines: routing, policy checks, tool selection and grading.

The move responds to an expensive production pattern: calling a general-purpose model dozens of times for yes/no or multi-class answers. As covered in why bounded outputs are increasingly replacing prose-producing AI, free-form completions add little for classification while token cost stays.

How it works

Each request carries a model ID, a shared input and a question map of up to 1,000 items. The shared input can be text, images, audio, video or PDF. Google bills it once per request; each question adds only its own instruction and criteria tokens, with no thinking-token, transcription or OCR charges. Total context is 262,144 tokens across the shared input and all questions. Google reports median model compute of 50–100 ms and 150–250 ms end-to-end for inputs up to 16,000 tokens, with latency nearly flat from 1 to 100 questions.

The API defines three question types. noul returns P(yes), bounded between 0.01 and 0.99. choice evaluates 2 to 255 options and returns normalized probabilities for all of them. score rates the input on 2 to 255 ordered rubric levels and returns a probability-weighted score. Every answer also carries a confidence value from 0 to 1; Google suggests routing low-confidence responses to a fallback or human reviewer. Questions are scored in parallel and cannot see one another.

All figures below are Google-reported. Google counts refusals and safety blocks as incorrect.

Google reports Gemini Decision 1 leads 7 of the 9 accuracy benchmarks. The calibration gap matters most: a Top-1 ECE of 0.0250, about half of Jev 1.14's 0.0494, means the probabilities track actual frequencies more closely and support more reliable automation thresholds. The model trails Jev 1.14 on When2Call (79.9% vs 82.0%) and posts its lowest score of 58.7% Macro-F1 on Amazon ESCI. On the shared suite, Google reports $0.88 versus $4.14 for Gemini 3.5 Flash-Lite; Jev 1.14 at $0.68 is cheaper.

Pricing

At $0.10 per million input tokens and $0 output, Gemini Decision 1 is below every general-purpose model in AI Mastery's pricing data as of October 22, 2026, including Gemini 3.1 Flash-Lite at $0.25 input and Mistral Small 4 at $0.15 input. The closer comparison is TypeSafe Jev 1.14, which MarkTechPost reports lists $0.042 per million input tokens, less than half Google's rate. Where calibration determines fallback frequency, the roughly 2× calibration advantage may justify the 2.4× input premium. For looser confidence requirements, Jev's lower price wins the shared suite by $0.20.

The endpoint is live on Gemini Developer API v1beta and requires google-genai SDK 3.2 or later, with a no-cost tier in Google AI Studio. MarkTechPost reports ecosystem integrations rolling out across OpenRouter, LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode; Gemini on Google Cloud is listed as coming soon. Google says internal use includes Gemini Notebook routing queries to model tiers with up to 40% lower latency, Google Labs' CC picking emoji reactions from more than 200 options, and Stitch's rapid canvas building playable wireframes in seconds.

AI Mastery analysis

The first limitation likely to bite is the absence of context caching and Batch API support. With the 262,144-token context budget, pipelines that reuse a large document or image corpus across requests pay $0.10 per million tokens in full on every call. OpenAI Decisions with GPT-6 Luna supports prompt caching at Luna's standard cached rates, which matters for high-volume retrieval and evaluation flows.

The choice question type also spreads probability across supplied options even when none fits. Google's documentation suggests adding an explicit "other" option or a yes/no scope guard. That consumes question slots and instruction tokens, so open-domain routing demands prompt engineering the flat API surface does not enforce.

The broader signal is that specialized inference endpoints are being carved off from generalist models at the infrastructure layer. Gemini Decision 1 cannot produce prose at all; it is sold on the accuracy of the numbers it returns. Whether its calibration lead holds on domain-specific benchmarks outside Google's published suite is the first question production teams should answer.

Sources

• Reported by: MarkTechPost: Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens
• Linked in that report: https://aistudio.google.com/docs/models/gemini-decision-1
• Competitor prices: AI Mastery token calculator, figures as of 2026-10-22