Gemini Decisions API: Google joins the labs selling decide, not write
Count them: OpenAI's Decisions API, Microsoft-Decision-1, Amazon's Strands Decider case study on Monday, and now Google. On Wednesday Google launched the Gemini Decisions API, powered by a new model called Gemini Decision 1 (gemini-decision-1). It does not write text. You hand it one input, text, images, audio, video or PDFs up to 256K tokens, plus up to 1,000 questions, and it scores the candidate answers in a single forward pass. Announcement: https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/, docs at https://aistudio.google.com/docs/decisions-api.
Three question types: yes/no probability, multiple choice from 2 to 255 options with normalized probabilities, and ordered rubric scores. Every answer carries a confidence between 0 and 1, so your code can route the unsure cases to a bigger model or a human. Google's numbers: 150 to 250 ms end to end, 50 to 100 ms of that model compute, nearly flat from 1 to 100 questions per request for contexts up to 16K. Price is $0.10 per million input tokens and zero for output, and the shared input is billed once no matter how many questions you stack on it. Google says that comes out nearly 5x cheaper than Gemini 3.5 Flash-Lite on its benchmark suite and claims the lowest calibration error, on benchmarks it names only by category.
The internal users tell you where this sits in an agent. Gemini Notebook routes every query to a model tier and thinking level with it, at up to 40% lower latency. And Google Search uses it as a safety layer for MCP tools, catching adversarial behavior inside the latency budget. That is the pattern: tool selection, policy checks, routing, context grading. All the little closed choices an agent makes between real steps, which today get done by a full generative call that writes paragraphs to output one word.
Vendor charts are vendor charts, so the useful companion is MMDM (https://arxiv.org/abs/2611.14309), a benchmark posted the day before: 5,448 fresh questions across 16 task categories, option spaces from 2 to 1,024, inputs up to 64,817 tokens, tested on 28 decision models. Its finding is the honest caveat. Decision models do fine on common knowledge and fall apart when options depend on each other: Perplexity's Decider-V1.2 hits 78.75% on hard single-select world knowledge but 53.95% on option logic and 47.95% on constraint planning. So use these models for the switchboard, not the reasoning. Four big vendors now agree that the routing step of the agent loop deserves its own cheap, priced primitive. The open question is who publishes an MMDM score first.
← Back to all articles
Three question types: yes/no probability, multiple choice from 2 to 255 options with normalized probabilities, and ordered rubric scores. Every answer carries a confidence between 0 and 1, so your code can route the unsure cases to a bigger model or a human. Google's numbers: 150 to 250 ms end to end, 50 to 100 ms of that model compute, nearly flat from 1 to 100 questions per request for contexts up to 16K. Price is $0.10 per million input tokens and zero for output, and the shared input is billed once no matter how many questions you stack on it. Google says that comes out nearly 5x cheaper than Gemini 3.5 Flash-Lite on its benchmark suite and claims the lowest calibration error, on benchmarks it names only by category.
The internal users tell you where this sits in an agent. Gemini Notebook routes every query to a model tier and thinking level with it, at up to 40% lower latency. And Google Search uses it as a safety layer for MCP tools, catching adversarial behavior inside the latency budget. That is the pattern: tool selection, policy checks, routing, context grading. All the little closed choices an agent makes between real steps, which today get done by a full generative call that writes paragraphs to output one word.
Vendor charts are vendor charts, so the useful companion is MMDM (https://arxiv.org/abs/2611.14309), a benchmark posted the day before: 5,448 fresh questions across 16 task categories, option spaces from 2 to 1,024, inputs up to 64,817 tokens, tested on 28 decision models. Its finding is the honest caveat. Decision models do fine on common knowledge and fall apart when options depend on each other: Perplexity's Decider-V1.2 hits 78.75% on hard single-select world knowledge but 53.95% on option logic and 47.95% on constraint planning. So use these models for the switchboard, not the reasoning. Four big vendors now agree that the routing step of the agent loop deserves its own cheap, priced primitive. The open question is who publishes an MMDM score first.
Comments