Google's Gemini Decisions API Scores 1,000 Questions Per Call

Google will let developers evaluate up to 1,000 questions in a single API call, at $0.10 per million input tokens and no output-token charges. The Gemini Decisions API, announced this week, arrives several weeks after OpenAI shipped a competing service and stakes a straightforward claim: routine application decisions do not need a generated paragraph to justify them.

The API runs on a new model, Gemini Decision 1. Instead of producing text token by token, it assigns probabilities to options a developer specifies. Applications can check whether a condition holds, pick from a predefined set, or score inputs against a rubric. In short, it answers rather than writes.

What can it actually process?

The service accepts multimodal input — images, audio and video — within a context window of up to 256,000 tokens. Google reported end-to-end latency of 150 to 250 milliseconds in tests on contexts of up to 16,000 tokens, with performance holding relatively stable from one question to 100.

Billing is the more interesting part. "Your shared input is billed once per request no matter how many questions you evaluate (each question adds only its own instruction and criteria tokens)," Google said in its announcement. That structure rewards the batch pattern the API encourages.

Confidence scores serve as the safety valve. Developers can use them to decide when an application proceeds automatically and when it refers uncertain cases to a human.

Where is Google using it already?

Google has deployed the technology internally in two places. In Gemini Notebook, it routes queries to different AI models based on request complexity, cutting latency for everyday queries by as much as 40%. In Search, it detects adversarial behavior involving Model Context Protocol (MCP) tools — a security use, not just a routing one.

The API is available now through Google AI Studio, with framework integrations rolling out. Gemini support on Google Cloud is promised later; Google has not named a date.

How does it compare with OpenAI?

OpenAI's Decisions API, generally available since October, caps at 400 questions per request against Google's 1,000. Both charge $0.10 per million input tokens at standard rates, with no output-token fees. OpenAI adds discounted pricing for cached inputs, which matters for applications that repeatedly process the same context. Google supports a broader range of multimodal inputs, including audio and video.

Omdia chief analyst Lian Jye Su plays down the head-to-head. "The model is great at generating decisions from unstructured, dynamic input," he said, naming policy and risk assessment and credit scoring as candidate workloads. He argues developers should look past published benchmarks and test how reliably each system decides under actual operating conditions. "The true benefit lies in the reduced misjudgment and improved decision-making quality," Su said.

What are the production risks?

Su warns that changes in incoming data, or deliberately manipulated inputs, can push decision APIs toward wrong judgments. A confidence score alone does not establish that an assessment is trustworthy. "All decision-making processes must be highly contextual with support from a trusted and secure context layer," he said.

His prescription: connect decision models to databases holding historical transactions and established business policies, so evaluations rest on organizational information rather than the immediate input alone. He also flags that Google must meet enterprise requirements for data retention and sovereignty without eroding the service's cost and performance advantages.

Tulika Sheel, senior vice president at Kadience International, raises the governance question. "As AI agents become more autonomous, enterprises will need clear accountability for decisions and actions taken on behalf of users," she said.

So what does this change for architects? A cheap, fast, probability-scoring layer now sits between your application logic and your generative models. Whether it earns trust is a production question, not a benchmark one.