Google has introduced the Gemini Decisions API, a new endpoint for applications that need bounded answers rather than generated prose. Released on November 11, the API can evaluate as many as 1,000 yes-or-no, multiple-choice or rubric-scoring questions in one request.

The endpoint accepts text, images, audio, video and PDF inputs within a 256,000-token input limit. Instead of returning an essay or chat response, it produces probabilities and confidence scores that software can use for classification, ranking, policy checks and routing. Google says requests can complete with end-to-end latency of 150 to 250 milliseconds, although that performance figure has not yet been independently established across production workloads.

A separate API surface for decisions

The release is more consequential than adding another Gemini model to a provider catalog. Most OpenAI-compatible integrations assume a generative request: the client sends messages or input items, the model returns text or structured output, and the provider charges for some combination of input and output tokens.

Gemini Decisions uses a different contract. An application supplies source material and a bounded set of questions, then receives numerical evidence for each possible answer. That design is suited to tasks such as deciding whether content should enter a review queue, selecting a model for a request, grading retrieved context or applying a documented scoring rubric.

Developers could approximate these workflows with ordinary text generation, but generated answers require parsing and can introduce wording that the application does not need. A purpose-built decision endpoint provides a more explicit interface: the meaningful output is the selected option and its associated probability or confidence, not a passage of prose.

Pricing changes the metering model

Google lists the price at $0.10 per million input tokens and does not bill output tokens. That is a substantial departure from the input-output accounting used by conventional generation models. It also means a request may return commercially important data even though its billable output-token count is zero.

Gateway operators therefore cannot rely on token totals alone. Usage records need to preserve the workload type, input volume, number of questions and relevant response metadata. Systems offering unified AI API billing may also need a separate rate-card rule so that a zero output-token value is not mistaken for a missing or incomplete usage record.

Batching creates another operational question. Supporting up to 1,000 questions in one request can reduce network overhead, but businesses will still need to decide how that work is attributed internally. One API call might represent hundreds of moderation checks, routing decisions or document scores performed for different products or customers.

Compatibility goes beyond changing a model name

The model is available through the Gemini API in Google AI Studio. For teams calling Google directly, adoption will require a client that understands its decision-oriented request and response structures. For a multi-model API gateway for teams, exposing the endpoint as if it were an ordinary chat model would hide important behavior and could make error handling, analytics and billing ambiguous.

A cleaner gateway design would identify decisions as a distinct workload class. Routing policy could then send bounded classification and scoring work to the new endpoint while reserving generative models for requests that actually require language output. Admission controls should also validate the question count, supported media and input size before forwarding a request.

The confidence data deserves careful treatment. A score is useful for thresholds and escalation rules, but it is not proof that a decision is correct or calibrated for every domain. Teams using the API for consequential policy or compliance workflows will need their own evaluations, threshold tests and human-review paths rather than adopting one universal cutoff.

What remains unproven

Google’s announcement establishes the API shape, published price and claimed latency, but independent evidence about production availability and performance remains limited. Real-world latency will also depend on payload size, media processing, geography and request volume.

The immediate engineering task is therefore controlled evaluation rather than broad replacement of existing classifiers. Developers should test representative multimodal inputs, compare confidence scores with known outcomes and verify how errors and partial results behave in large question batches. Gateway and platform teams should additionally confirm that their logs can distinguish a decision request from generation before making the endpoint available to customers.