Google’s new Gemini model makes decisions without writing answers

Google DeepMind introduced the API on November 11, 2026, in an announcement by product manager Min Choi and principal scientist Felix Yu. Powered by Gemini Decision 1, it can, Google says, evaluate up to 1,000 independent yes-or-no, multiple-choice and rubric-scoring questions about shared material in one request. This is a connection for other software to use, not another chatbot interface.⁠2

The distinction matters. A conventional generative model builds an answer token by token; an application seeking a decision may then have to extract it from that answer. Gemini Decision 1 instead evaluates bounded alternatives in a single processing pass and returns structured judgments and probabilities. Google’s documentation says software can branch, sort and route directly on the response, with output-token usage always zero. It does not supply an open-ended explanation.⁠1⁠3

That makes the endpoint potentially useful wherever the desired answer is an instruction to the surrounding application: choose a model, check a policy, rank relevant material or decide where a request should go. The engineering attraction is straightforward. If software needs a bounded judgment, making a model write a paragraph first is an avoidable detour.⁠1⁠3

The launch coverage in Hybrid.co.id describes three question types that developers can mix in one call. Yes-or-no questions estimate whether a condition holds, with optional descriptions of what counts as true or false. Multiple-choice questions accept two to 255 candidates and return probabilities normalized across those alternatives. Rubric questions use ordered levels beginning at zero, returning a continuous weighted score alongside probabilities for the individual levels.⁠1

Those definitions put considerable responsibility on the developer. Someone still has to choose the alternatives, define the scoring levels and decide what action follows. A well-formed response cannot repair a badly framed question. Nor does selecting the most probable option establish that the application offered the right options in the first place.

Consider the insurance example described in the launch coverage: a claim PDF can be supplied alongside photographs of damage. An application could ask bounded questions about that material and use the probabilities to determine the next step. That is a plausible intake mechanism, not evidence that the model can safely approve or reject claims. The distinction between routing a file and deciding someone’s entitlement should not disappear merely because both can be represented as a classification.⁠1

The advertised inputs include text, images, audio, video and PDFs, combined within the same request. Hybrid reports a context window of up to 256,000 tokens—the units the system processes—and adjustable image or video resolution. Lower resolution can reduce token consumption at the expense of visual detail. A customer-call recording, for example, can travel with account information rather than being handled as an entirely separate input.⁠1

The thousand-question claim also needs its proper meaning: many questions applied to shared material, not a thousand unrelated, full-size input packages processed at no extra cost. Each question brings its own instructions and criteria.⁠1

Hybrid reports median model-computation time of 50–100 milliseconds and end-to-end latency of 150–250 milliseconds for contexts up to 16,000 tokens. It also describes response duration as staying nearly steady between one and 100 questions. These are narrower claims than the headline capacities: they do not establish the same speed at 1,000 questions or at the largest advertised context.⁠1

The speed figures have not been independently reproduced in the reporting available for this launch. Nor is there a reproducible account of hardware, serving region, concurrency, warm-up conditions or the full distribution of response times. An application should treat them as a reason to run a test, not as a service-level guarantee.⁠20

The reported price is $0.10 per million input tokens, with no output-token charge. Shared material is described as billed once within a request, however many questions use it; the questions’ instructions and criteria still add input tokens. Hybrid also describes no variable thinking-token charges or separate audio-transcription and optical-character-recognition fees. The service is therefore not free to process material. Its zero-output distinction is more fundamental than a discount: there is no generated answer text to bill.⁠1⁠3

A launch comparison calls the endpoint “almost five times” cheaper than Gemini 3.5 Flash-Lite through the Interactions API in equivalent benchmark testing. Without a reproduced methodology, that is not a universal saving. Actual costs depend on the material submitted, the number and length of questions, and the workload being replaced.⁠1

Google describes the returned probabilities as calibrated: broadly, that means predicted likelihoods should track how often outcomes are actually correct. Hybrid describes confidence information on a zero-to-one scale, which developers can use to set thresholds for fallback procedures or human review. A high value remains a prediction, not a guarantee.⁠1⁠3

The launch material promotes results in routing, reasoning, ranking and multimodal understanding, including a lowest-calibration-error claim. It does not establish independently validated calibration measurements, a general error rate or a reproducible accuracy comparison. Those omissions matter most where an incorrect classification triggers an action that is difficult to reverse.⁠1⁠20

Google’s described internal uses show the intended role. A workflow called Gemini Notebook selects model tiers and thinking levels, reserving deeper reasoning for more complex queries; it carries a claimed latency reduction of up to 40 percent for everyday queries. That claim belongs to that workflow, not every application using the endpoint. In Google Labs’ CC application, the task is smaller: choosing an emoji from more than 200 options in a fraction of a second.⁠1

Other examples need the same separation between the decision model and the surrounding product. Stitch is described as using it within rapid canvas experiments involving interactive wireframes and voice- or touch-driven revisions. A Google Search example checks Model Context Protocol tools—connections between AI systems and tools—for adversarial behavior and harmful content. The safety example supplies no numerical detection accuracy or false-positive rate. It illustrates an intended use, not demonstrated immunity to attacks.⁠1

Access is described through the Gemini API in Google AI Studio using an API key, with Python, JavaScript and REST guides and a no-cost prototyping tier. The model identifier is gemini-decision-1. Announced ecosystem support includes the google-genai software kit and integrations involving OpenRouter, LangChain, Vercel and other frameworks. Those are launch support claims, not verified customer production results. Google Cloud availability is described as forthcoming, without a firm date.⁠1⁠2

Country eligibility and preview-versus-general-availability status remain unclear in the launch reporting. More importantly, the consequential test is still ahead: whether speed, cost and confidence quality hold on developers’ own workloads. A model that never writes an answer can be valuable. Software still needs a sound reason to act on the answer it implies.