Google launches Gemini Decisions API for fast, cheap multimodal AI choices

When a product needs to decide among dozens of options—like routing a user request, grading a form, or picking an emoji—developers traditionally call a generative model, wait for token‑by‑token output, and pay for every generated word. Google’s Gemini Decisions API flips that model on its head, offering a purpose‑built endpoint that returns a probability for each candidate in a single, fast pass.

The new service bundles up to 1,000 closed‑set questions—yes/no, multiple‑choice, or ordered rubric—against a single multimodal input that can include text, images, audio, video or PDFs. By scoring all options together, the API eliminates the need for separate classifiers or repeated calls, delivering decisions in milliseconds while keeping the cost per request low. This approach is especially useful for real‑time UI feedback where milliseconds matter.

Source: Google AI Blog: Gemini Decisions API: Fast, calibrated multimodal decisions ↗

According to Google, the Decisions API can accept one shared multimodal input and up to 1,000 independent questions, returning a confidence score from 0 to 1 for each. The response includes the winning choice for multiple‑choice queries and a weighted score for rubric evaluations, letting downstream code route ambiguous cases to a human reviewer. The API also returns the raw probability distribution, enabling fine‑grained ranking beyond the top choice.

The underlying Gemini Decision 1 model scores candidates in a single forward pass, delivering probabilities and confidence values in 150–250 ms end‑to‑end latency. Google prices the service at $0.10 per 1 M input tokens with zero output‑token charges, making it roughly five times cheaper than the comparable Gemini 3.5 Flash‑Lite interaction endpoint while achieving the lowest calibration error on benchmark suites. These latency figures include the full network round‑trip, making the service suitable for interactive applications.

Source: Google AI Blog: Gemini Decisions API: Fast, calibrated multimodal decisions ↗

The API is limited to closed‑set decision formats, so it cannot generate free‑form text or handle open‑ended reasoning tasks. Developers must translate their problem into yes/no, choice or rubric prompts and supply appropriate criteria, which can add engineering overhead for use cases that were previously handled by a single generative call. Moreover, the confidence scores are only as reliable as the training data, so edge cases may still produce misleading probabilities.

While Google lists integrations with LangChain, Vercel and other SDKs, the ecosystem is still rolling out, and the service currently runs only in Google AI Studio and select demo apps. Large‑scale production deployments will need to monitor latency under load and may face throttling or quota limits until broader cloud‑wide support arrives. Early adopters should also plan for version upgrades, as Google may introduce new model variants that change pricing or latency.

Developers can start experimenting today by grabbing an API key in Google AI Studio and following the Python, JavaScript or REST guides that Google provides. The blog post includes sample code that sends a mixed image‑text request and three different question types in a single call, plus demo apps like Doorstep Parcels and Speed Sketch to see the latency in action.

A sensible first project is a low‑cost routing layer—e.g., directing support tickets to the right department or flagging policy‑violating content. By checking the confidence score returned for each decision, teams can automatically handle high‑certainty cases and fall back to a human reviewer when the score dips below a chosen threshold, balancing speed with safety. Monitoring tools in Google AI Studio let you track per‑request latency and token usage, helping you tune thresholds.