Google Launches Gemini Decisions API to Score 1,000 Questions Per API Call

Quick Facts

• The Gemini Decisions API, powered by the new Gemini Decision 1 model, launched November 11, 2026, and is available now in Google AI Studio.
• A single API request can evaluate up to 1,000 questions against one multimodal input, returning confidence scores in 150-250 ms at $0.10 per million input tokens with no output tokens billed.
• An independent benchmark places Gemini Decision 1 ninth of 18 decision models at launch, with higher costs than the category median.

Google launched the Gemini Decisions API on November 11, 2026, targeting a specific slice of AI workload: bounded decisions inside applications and agents. The API is built around Gemini Decision 1, a purpose-built model that scores candidate options in a single forward pass rather than generating text token by token.

The endpoint handles yes/no checks, multiple-choice questions from 2 to 255 options, and ordered rubric scoring. Every answer carries a confidence score from 0 to 1. Developers access it through the Gemini API using the model ID gemini-decision-1.

How It Works

Each request accepts one multimodal input spanning text, images, audio, video, and PDFs, up to 256,000 tokens. Against that input, developers can run up to 1,000 independent questions. Google DeepMind Product Manager Min Choi and Principal Scientist Felix Yu wrote that median model compute runs 50 to 100 ms and stays nearly flat from 1 to 100 questions per request on inputs up to 16,000 tokens.

The API does not support context caching, the Batch API, or Flex and Priority inference at launch. Google has not published latency or accuracy figures for requests that reach the full 1,000-question limit.

Google Already Uses It in Production

Google says it deployed the API internally before the public launch. Gemini Notebook uses it to route queries to a model tier and thinking level, which Google says cuts latency by up to 40% on everyday requests. Google Search uses it as a safety layer for MCP tools to detect adversarial behavior. The Stitch design tool uses it to generate playable wireframes in seconds, and Google Labs CC uses it to pick one emoji reaction from more than 200 options.

Pricing and Competitive Position

At $0.10 per million input tokens with zero output tokens billed, the pricing model is straightforward. Running Gemini Decision 1 on the independent Artificial Analysis Decision Index costs an average of $345 per million tasks, against a category median of $143.

Speed tells a similar story. Gemini Decision 1 posts a Time per Task of 211 ms on the Decision Index, faster than Drex Max v1.1 at 283 ms and GLiDE v1.2 at 284 ms, but slower than Celeris-1.1 Decision at 61 ms, Elephant 9B at 79 ms, and Jev 1.14 at 92 ms.

Benchmark Results Are Mixed

The independent Artificial Analysis Decision Index v2.1 ranks Gemini Decision 1 ninth of 18 models at launch. The index combines six evaluations covering document review, tool calling, knowledge, and reasoning. Google's model scores highest among decision models on the knowledge and reasoning portions of that index.

Google's own benchmarks, which cover agent routing, reasoning, ranking, and multimodal understanding, show Gemini Decision 1 leading in seven of nine tests against Jev and OpenAI's Decisions API, with the lowest calibration error. Those figures come from Google, not an independent evaluator.

What This Means for Builders

The API targets workloads where full text generation is wasteful: routing requests to the right model, running policy checks, grading retrieved context, or selecting tools inside an agent loop. For teams running millions of such decisions daily, eliminating output token costs matters more than raw benchmark rank.

The tradeoff is cost per task. At more than twice the category median price on independent tests, teams with high-volume, latency-sensitive workloads will need to benchmark their own traffic before committing. The API is available now through Google AI Studio with no waitlist.