Google's Gemini Decisions API Pushes Bounded Choice at Scale

Each question type has a defined output. Yes/no items estimate the probability that a condition holds, choice items return a winner and normalized probabilities for every candidate, and rubric items return a weighted score on a continuous scale plus probabilities for each level. Every answer carries a 0-to-1 confidence value so code can send uncertain cases to a fallback or human reviewer, according to Google. [1]

The common context incurs one charge per call, Google says, while each question adds only its own instruction and criteria tokens. The input uses the Content block format already used by the Interactions API, and Google's example tests a breakroom photo plus a text note against all three question types in a single call. [1]

The product's significance lies in its narrowness. Rather than asking a model for free-form reasoning, the Decisions API returns bounded selections and scores. Google contrasts this with token-by-token decoding, which can introduce variable output and thinking-token costs in generative calls. [1]

Google says the model was built on Gemini's native multimodal architecture and ingests text, images, audio, video and PDFs in one forward pass. The intended benefit is that a single shared context can answer many bounded questions without separate transcription or document-extraction steps. [1]

That pattern fits agent workflows that must make many small judgments: which tool to call, whether a policy applies, how to grade retrieved context or which model tier should handle a query. Google says Gemini Notebook uses the endpoint as a real-time router, classifying intent in milliseconds and reducing latency by up to 40% on routine queries while sending harder work to deeper reasoning models. [1]

Google also cites Stitch using the API to create playable wireframes and apply voice or tap edits almost instantly; CC from Google Labs using it to choose an emoji from more than 200 options; and Google Search using it as a safety layer for MCP tools, detecting adversarial behavior and content within its latency target. [1]

Google presents the cost case as a comparison. It pits Gemini Decision 1 on the Decisions API against Gemini 3.5 Flash-Lite on the Interactions API, saying the newer setup is almost five times less expensive over a common benchmark set. It also says there are no separate fees for output words, variable thinking, audio transcription or OCR. [1]

Google also claims benchmark leadership in agent routing, reasoning, ranking and multimodal understanding, along with the lowest calibration error in its suite. The post does not describe the suite's size or composition, so the numbers should be read as Google's launch claims. [1]

The announcement defines itself against token-by-token generation. Google says the new model scores candidate options without decoding answer text, a design intended to keep latency low as the number of questions rises. [1]

The comparison model named in the post is Gemini 3.5 Flash-Lite paired with the Interactions API; Google uses it for the cost comparison. The new model sits inside the existing Gemini API and AI Studio rather than on a separate platform. [1]

Separately, Google's blog homepage promoted Gemini 4 Hydrogen for coding and agentic workflows, TPU 8tx and 8i Max for the agentic era, and Gemini features across Chrome and Pixel. Those homepage items are distinct from the Nov. 11, 2026 Decisions API developer post, which describes the new endpoint and its model. [1] [2]

According to Google, its own product groups rely on the endpoint for quick operational choices; they add capabilities by specifying candidate choices, not by building bespoke classifiers or holding requests for a generative response. That description is an internal-adoption claim, not a customer case study. [1]

Google says integrations are rolling out with OpenRouter, LangChain, LiteLLM, Vercel, Pydantic AI, LiveKit, Pipecat and OpenCode, and that Gemini in Google Cloud support is coming soon. A no-cost tier in AI Studio is available for prototyping, and the common context incurs one charge per call however many questions are attached. [1]

The announcement also points developers to API-key access in Google AI Studio and to Python, JavaScript and REST guides. It names three demo apps — Doorstep Parcels, Speed Sketch and Find the Zoo Keeper — as ways to try the endpoint. Those materials accompany the launch; they do not establish independent performance, show how the API behaves outside Google's examples or validate its confidence scores on customer data. [1]

For developers, the appeal is bypassing task-specific classifier training. A team can describe candidate options and receive probabilities, which may shorten iteration. The tradeoff is dependence on a hosted Google model and on how well its confidence scores behave in the team's domain. [1]

For Google, the launch argues that not every AI task needs a generated answer. Some need selection, checking and scoring. If the latency and price claims hold in production, the endpoint could become a control layer inside larger agent systems, handling many small judgments around a generative core. [1]

The internal examples span consumer and developer surfaces, from Notebook and Stitch to CC and Search. That range suggests Google is testing the API in varied workflows, but the post discloses no usage volumes, error rates or production constraints. [1]

Pricing may alter how builders think about question volume. Because input is charged once and output is free, adding questions to a request has a different cost profile from repeated generative calls. Each question still adds instruction and criteria tokens, and the total depends on input size and modality, Google says. [1]

Google also says there are no separate audio transcription or OCR charges, simplifying cost estimates for mixed-media work. The shared-input model means one large context can serve many questions without multiplying the input charge, according to the announcement. [1]

The main open question is independent performance. Google reports benchmark leadership and low calibration error, but the post gives no suite size or full protocol. Outside tests will be needed to judge how the model performs on real routing and policy tasks. [1]

Latency scaling is another gap. Google's median figures cover contexts up to 16K tokens, while the API accepts inputs up to 256K. The post does not say how latency changes at the upper end or how pricing behaves for very large multimodal inputs. [1]

The announcement also leaves open how the service handles ambiguous or adversarial inputs beyond the Search safety example, and whether confidence values are comparable across modalities. [1]

Mixing question types raises a further issue. Choice items return normalized probabilities, while rubric items return per-level probabilities and a weighted score. The post does not specify whether those outputs are directly comparable when a system sets fallback thresholds. [1]

Enterprise controls are unspecified. Google says Gemini in Google Cloud support is coming and integrations are rolling out, but the post does not detail data residency, retention, audit logs or access controls. Those details will determine whether regulated customers can adopt the endpoint. The post also does not address customer-supplied evaluation data or fine-tuning. [1]

Confidence scores are only useful if thresholds are tuned. Google says answers include a 0-to-1 score and that uncertain cases can go to a fallback or human reviewer, but it publishes no recommended thresholds or domain-specific calibration guidance. [1]

The next signals are outside benchmarks, named customer deployments and the Gemini in Google Cloud release. Google points developers to Python, JavaScript and REST guides plus demo apps including Find the Zoo Keeper, Speed Sketch and Doorstep Parcels. Adoption patterns in those tools may show which use cases fit. [1]

Ecosystem shipping is another signal. Google named OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode as integrations rolling out, but it gave no dates for each. The list is a watch item, not a guarantee of availability. [1]

The larger question is whether bounded decision endpoints become a common layer in agent stacks. Google's post makes that case; it does not establish the outcome. The answer will depend on measured accuracy, calibration and cost in production, not launch-day claims alone. [1]

Sources

Google AI announcements

blog.google

Prepared with AI assistance from the cited sources.