Google Built an AI Model That Never Writes a Sentence, and That's the Point
AI & ML

Google Built an AI Model That Never Writes a Sentence, and That's the Point

The Gemini Decisions API scores up to 1,000 questions against a single input without generating any text, cutting cost and latency for the routing, scoring, and compliance checks hiding inside every agentic workflow. It is a quiet admission that most AI decisions in production do not need an essay.

PublishedNovember 19, 2026
Read time6 min read
Share

What a decision model actually does differently

A conventional generative model, whatever its size, produces output one token at a time, assembling a sentence or paragraph through a sequential prediction process designed for open-ended text. The Gemini Decisions API, running on a new model called Gemini Decision 1, skips that process entirely for a specific category of task: it assigns calibrated probabilities to a predefined set of options you supply, whether that means checking whether a condition holds true, selecting one choice from a fixed list, or scoring an input against a rubric you define in advance.

The practical difference shows up immediately in what the API can do in a single call. Developers can submit up to 1,000 separate questions against one shared input, and the model accepts images, audio, and video alongside text, with a context window reaching 256,000 tokens. An application built on top of it can use the returned confidence scores to decide automatically whether to proceed or route the uncertain cases to a human reviewer, turning what used to require a chain of separate model calls into one scored batch response.

The economics that make this worth re-architecting a workflow around

Pricing is set at 0.10 dollars per million input tokens with no output token charge at all, and Google's billing model treats shared input as a single cost regardless of how many questions get evaluated against it: run 20 separate checks against one 4,000-token record and you pay roughly 0.50 dollars per 1,000 records, against roughly 8.10 dollars for making 20 separate standard API calls against the same record. That is not an incremental optimization. It is a pricing structure built specifically to reward exactly the batched, repetitive scoring pattern that agentic workflows generate constantly and that most enterprises have been paying full generative-model rates to run.

Latency tells a similar story. Google reports end-to-end response times of 150 to 250 milliseconds for contexts up to 16,000 tokens, and that latency holds relatively steady whether the request contains one question or 100, since the model is scoring predefined options rather than generating variable-length text. For any workflow where a routing decision, a compliance check, or a content classification currently sits in the critical path of a user-facing agent, that latency profile is the difference between a decision step that is invisible to the end user and one that visibly slows the interaction down.

Where Google is already using its own product

Google is not launching this as a speculative capability; it is already running Gemini Decision 1 inside its own products. Gemini Notebook uses it to route incoming queries to different underlying models based on how complex each query actually is, and Google says that routing cuts latency for everyday queries by as much as 40 percent compared to sending every query through a single, larger model regardless of complexity. That is a direct, measurable validation of the core thesis behind decision models: a large fraction of the judgment calls inside a production AI system do not require the model doing the heavy generative lifting, they require a fast, cheap, accurate classifier sitting in front of it.

Google is also deploying the same underlying model inside Search to detect adversarial behavior targeting Model Context Protocol tools, a genuinely new category of security use case that did not exist as a product requirement before agentic systems started exposing MCP interfaces broadly. Analyst Lian Jye Su of Omdia points to policy and risk assessment and credit scoring as additional use cases enterprises should expect to see decision models applied to, domains where a calibrated probability score with an audit trail is often more defensible than a generated explanation.

What a competing OpenAI product one month earlier tells you about this category

OpenAI shipped its own Decisions API in October, roughly a month before Google's launch, supporting up to 400 questions per request without audio or video input support. Google's version launched supporting 2.5 times the batch size with broader modality coverage already built in, which suggests this specific capability moved quickly from a single vendor's differentiator to a category every major model provider now feels compelled to field a competitive version of. That pattern, where a genuinely new API shape gets matched by a rival lab within weeks rather than quarters, is itself useful signal about how fast the competitive cycle has compressed around anything that demonstrably reduces enterprise AI spend.

Omdia's Su also raises the caveat that matters most before any enterprise re-architects a workflow around either vendor's decision API: a confidence score alone does not prove an assessment is trustworthy, and shifts in incoming data, or inputs deliberately crafted to manipulate the scoring, can produce confidently wrong judgments. His recommendation, connecting decision models to historical transaction data and documented business policy rather than treating the raw confidence score as sufficient justification on its own, is the right starting discipline for any team evaluating either Google's or OpenAI's version for a production decision that actually matters.

What this means for your own agentic architecture

Audit your current agentic workflows specifically for places where a full generative call is answering what is actually a closed-set question: approve or deny, route to A or B, score this input against five criteria. Those are the exact patterns the Gemini Decisions API and its OpenAI equivalent were built to intercept at a fraction of the cost and latency of a standard call, and if your architecture is currently running every one of those checks through a general-purpose model, you are very likely overpaying for a decision your workflow does not need generated prose to make.

Before committing a compliance-sensitive decision to either API, build in the accountability layer Kadence International's Tulika Sheel and Omdia's Su both flagged: a documented record of which policy or historical data informed each automated decision, not just the confidence score the model returned. Tying agent AI spend to actual decision quality, the same principle behind Splunk's newly launched Tokenomics tooling, matters just as much here, since a decision model that is cheap per call but wrong often enough to trigger downstream rework has not actually saved your organization anything.

Tagged#news#ai-ml#ai#llm#agents#agentic-ai#openai#anthropic#regulation#gemini-decisions-api#gemini-decision-1#google-deepmind#model-context-protocol#decision-models#ai-cost-optimization#model-routing