Google launches Gemini Decisions API for fast, text-free AI choices
Google released the Gemini Decisions API on November 12, 2026, enabling low-latency, low-cost evaluation of up to 1,000 closed-set options without generating text.
On November 12, 2026, Google introduced the Gemini Decisions API, a specialized tool designed to handle routine logical choices in AI applications without generating natural language text. This new service, powered by the Gemini Decision 1 model, aims to reduce latency and costs for developers who need fast, deterministic outcomes rather than creative prose.
What happened
The Gemini Decisions API allows software systems to evaluate predefined options in a single pass, bypassing the computational overhead of traditional generative models. Instead of producing a sentence or paragraph, the model scores up to 1,000 closed-set choices against a given input. Google states that response times for tested workloads range from 150 to 250 milliseconds, significantly faster than standard large language model interactions. Pricing is structured to favor high-volume decision-making, starting at $0.10 per million input tokens with no charges for output tokens.
This release places Google in direct competition with other tech giants moving toward specialized decision models. OpenAI launched its own Decisions API just weeks prior, while companies like TypeSafe AI and Microsoft are also developing similar tools. The market shift reflects a growing recognition that many AI tasks do not require generative capabilities. Google CEO Sundar Pichai described the model as “built for closed-set choices,” emphasizing its role in handling bounded, specific decisions rather than open-ended generation.
Google is already deploying this technology internally. In Gemini Notebook, the model handles real-time routing for models and reasoning levels, reportedly reducing latency by up to 40%. Additionally, Google Search uses the system to detect potentially adversarial behavior involving Model Context Protocol (MCP) tools. These internal use cases demonstrate the model’s capacity to handle high-frequency, low-latency tasks that would be inefficient or expensive using general-purpose generative models.
How it works
The core mechanism of the Gemini Decisions API is its ability to process a shared context once and then evaluate multiple independent questions against it. The model supports a context window of 256,000 tokens and can handle multimodal content. Developers define a set of possible answers or actions, and the model assigns probabilities and confidence scores to each option. This approach eliminates the need for the model to generate text token-by-token, which is the primary source of latency and cost in traditional LLMs.
Because the API evaluates each question separately, it is best suited for parallelizable tasks where one decision does not depend on the outcome of another within the same call. If a workflow requires sequential logic—where answer A determines question B—developers must make separate API calls. However, the cost structure is optimized for batch processing: developers pay only once to send the shared context, even if they ask hundreds of questions about it. The questions themselves still count toward token usage, but the elimination of output token costs makes high-volume scoring economically viable.
Key details
- Model name: Gemini Decision 1, designed specifically for closed-set choices.
- Performance: Response times between 150 and 250 milliseconds for tested workloads.
- Capacity: Supports a 256,000-token context window and up to 1,000 independent questions per input.
- Pricing: Starts at $0.10 per million input tokens; no charges for output tokens.
- Benchmarks: Google claims a 92.2% score on JevBench, compared to 90.9% for OpenAI’s GPT-6 Luna and 90% for TypeSafe’s Jev 1.14.
- Availability: Accessible via Google AI Studio with a free prototyping tier, and through integrations like OpenRouter, LangChain, Vercel, and LiteLLM.
Why it matters
For software engineers building AI agents or automated workflows, the Gemini Decisions API offers a way to decouple reasoning from generation. Many applications spend significant resources generating text only to parse it back into a structured decision. By skipping the generation step, developers can achieve lower latency and more predictable costs. This is particularly valuable for real-time applications, such as fraud detection, dynamic routing, or interactive user interfaces, where delays of even a few hundred milliseconds can degrade the user experience.
The competitive landscape for AI infrastructure is shifting from general-purpose chatbots to specialized primitives. As noted in the benchmarks, while Google’s model scored higher on overall accuracy, TypeSafe’s Jev 1.14 was cheaper to run for the full benchmark suite ($0.68 versus $0.88). This suggests that developers will need to carefully evaluate trade-offs between accuracy, calibration error, and cost based on their specific use cases. The fact that all benchmark figures come from Google’s own testing means independent validation in production environments will be crucial for teams making long-term architectural decisions.
What you can do
- Audit your current AI workflows to identify steps that involve simple classification or selection from a fixed list of options.
- Prototype with the free tier in Google AI Studio to test how Gemini Decision 1 handles your specific closed-set questions.
- Compare latency and costs against your existing generative model setup, paying attention to the savings from eliminated output tokens.
- Evaluate integration options such as LangChain or LiteLLM to see how easily the API fits into your current stack.
- Monitor calibration errors in your tests, as Google highlights low calibration error as a key strength of Gemini Decision 1.
- Plan for sequential logic by designing your architecture to handle separate API calls when decisions depend on previous outcomes.



