Gemini Decisions API: fast, calibrated decisions from Google

What did Google announce?

On November 11, Google introduced the Gemini Decisions API. It is a new endpoint for closed-set decisions that runs on the new Gemini Decision 1 model (gemini-decision-1). It is available through the Gemini API in Google AI Studio (Google blog).

Google positions it for tool selection, policy checks, model routing and context grading. A single request takes one multimodal input of up to 256K tokens, plus up to 1,000 independent questions. Instead of generating an answer token by token, the model scores the given options in a single forward pass, then returns probabilities and confidence scores.

Three question types in one request

• Yes/no: the probability that a condition holds, with optional criteria.
• Multiple choice: picks from 2 to 255 options, with normalized probabilities for every option.
• Rubric scoring: on an ordered scale starting at 0, it returns a continuous weighted score and per-level probabilities.

Every answer comes with a confidence score between 0 and 1, so code can send ambiguous cases to a fallback or a human reviewer. The input can combine modalities, for example a call recording with account metadata, or a scanned PDF claim with photos of the damage.

What do the numbers show?

According to Google, median model compute is 50-100 ms, and end-to-end latency is 150-250 ms for multimodal inputs of up to 16K tokens. This stays nearly flat from 1 to 100 questions per request. On the company's own benchmark suite, Gemini Decision 1 leads in agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error.

Source: Google. Gemini Decision 1 against TypeSafe Jev 1.14, GPT-6 Luna (OpenAI Decisions API) and Gemini 3.5 Flash-Lite, Top-1 accuracy unless noted otherwise.

In the table it scores 92.2% on JevBench (GPT-6 Luna: 90.9%, Jev 1.14: 90.0%), and its Top-1 ECE calibration error is 0.0250 (GPT-6 Luna: 0.0721). It is not first everywhere: on Mind2Web Multimodal and When2Call, Jev 1.14 does better. Jev also has the lowest total price across the seven shared evals, $0.68, versus $0.88 for Gemini Decision 1. Google says that across the shared suite it is nearly 5x cheaper than Gemini 3.5 Flash-Lite on the Interactions API. These are Google's own measurements.

How much does it cost?

The price is $0.10 per 1M input tokens, and output tokens are free. The shared input is billed once per request no matter how many questions run on it, and each question adds only its own instruction and criteria tokens. There is no separate charge for thinking tokens, audio transcription or OCR. A no-cost tier is available in Google AI Studio for prototyping.

Where is Google already using it?

• Gemini Notebook: as a real-time router, it picks the model tier and thinking level. Google says this cuts latency by up to 40% for everyday queries.
• Stitch: the rapid canvas experiment builds playable wireframes of a product flow in seconds, and applies edits requested by voice or tap almost instantly.
• CC (Google Labs): it picks the right emoji reaction from more than 200 options in a fraction of a second.
• Google Search: as a safety layer, it detects adversarial behavior and content for MCP tools.

Integrations are rolling out to OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs. Support for Gemini in Google Cloud is coming soon, according to the company.

Why it matters

This is Google's answer to the OpenAI Decisions API, which became generally available on October 22 with GPT-6 Luna. Decision models keep arriving on the open side as well, for example d3-ultra, joey-1, Drex Max v1 and Sarvam Prism.

For anyone building agents, routing, tool selection and guardrail checks can become a fast, cheap separate call, without running a generative model or a custom classifier for them. All the published numbers come from Google, so it is worth measuring on your own data before switching.