Google launches Gemini Decisions API and Gemini Decision 1 model for fast, single-pass multimodal classification at $0.10 per 1M input tokens

Google has released a new Gemini API endpoint, client.decisions.create, and a model called gemini-decision-1 through Google AI Studio. It answers up to 1,000 yes/no, multiple-choice (2 to 255 options) or rubric-score questions about one shared input in a single pass. The input can be text, images, audio, video or PDFs, up to 256K tokens. It returns probabilities and confidence scores rather than generated text. Pricing is $0.10 per 1M input tokens with no output charge, plus a free tier. Google claims 150-250 ms latency. SDKs exist for Python, JavaScript and REST, and integrations with OpenRouter, LangChain, Vercel, LiteLLM and others are described as rolling out. Google Cloud support is "coming soon".

Teams that use a large model for routing, policy checks, grading or tool selection could swap in a cheaper, faster step. They would not need to train their own classifiers or pay for generated tokens. Asking many questions of one input in a single pass could cut cost and delay in moderation, triage and evaluation pipelines. The effect is real but narrow: it only helps where the task is a bounded decision, and Google Cloud customers cannot yet use it through their usual route. Because it is a Gemini-only endpoint, adopting it also deepens dependence on Google.

The claims of best-in-class benchmarks, the lowest calibration error and being "nearly 5x cheaper" than Gemini 3.5 Flash-Lite are Google's own, and no independent test supports them. The 5x figure compares against one specific setup (Flash-Lite with the Interactions API), so it may not hold against a tuned small classifier. Internal wins such as "up to 40% lower latency" in Gemini Notebook are best-case, unverified numbers. "Calibrated" probabilities also need testing on your own data, since good average calibration can still fail on rare or out-of-distribution inputs.

1. 01Independent benchmarks and developer reports of accuracy, calibration error and real end-to-end latency against Gemini Flash-Lite, small open-weight classifiers and fine-tuned models on real workloads, especially with video, audio and many questions per call.
2. 02Whether the partner integrations (OpenRouter, LangChain, LiteLLM, Vercel) actually ship and work in production, and when Gemini in Google Cloud and Vertex gets it, which matters most for enterprise adoption.
3. 03Pricing and limits after launch: how the free tier is capped, whether $0.10 per 1M input tokens holds, and whether rivals such as OpenAI, Anthropic or open-source providers launch comparable low-cost classification endpoints within the next few months.