What happened
Google’s announcement, dated November 11, 2026, introduces the Gemini Decisions API, powered by Gemini Decision 1, identified as gemini-decision-1. It is available through the Gemini API in Google AI Studio. Google describes it as an endpoint for closed-set decisions: instead of decoding an answer token by token, the model scores candidate options in a single forward pass and returns probabilities and confidence scores.
A request can contain one shared multimodal input spanning text, images, audio, video, and PDFs, up to 256K tokens, and as many as 1,000 independent questions. The API supports three question types: yes/no estimates, multiple choice with 2 to 255 choices, and ordered rubric scores. Each answer has a confidence score from 0 to 1; multiple-choice responses also return normalized probabilities for each candidate, while rubric scoring returns a weighted score and probabilities per level. Google says a developer can combine modalities in one input, such as a call recording and account metadata, or a scanned claim and damage photos. Image and video resolution can be tuned to control visual detail and token cost.
The listed price is $0.10 per 1 million input tokens and zero output-token charges. The shared input is billed once per request regardless of the number of questions, with question instructions and criteria adding their own tokens. Google says there is an API no-cost tier in AI Studio for prototyping, and that there are no separate audio transcription or OCR costs. Its sample Python request uses the google-genai client and the v1beta API version. Documentation is also listed for JavaScript and REST.
Google says the model has 50–100 ms median model-compute latency and 150–250 ms end-to-end latency, with latency remaining nearly flat from one to 100 questions per request. It reports that the model leads across a benchmark suite covering agent routing, reasoning, ranking, and multimodal understanding, and has the lowest calibration error. Google further claims nearly five times lower cost across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. It names internal uses: Gemini Notebook routes queries to model tiers and thinking levels; Stitch uses it for rapid design iteration; CC selects emoji; and Google Search uses the Decision API as a safety layer for MCP tools. Ecosystem integrations are rolling out, including OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode. Support in Gemini in Google Cloud is described as coming soon.
What the archive adds
New Runtime’s earlier coverage of Gemini 4 Carbon availability on Google Cloud and in Agent Studio (November 11, 2026) concerned developer access to a different Gemini model and platform surfaces. This is a new event, not a repeat of that model availability story: the new fact is a dedicated decision-scoring endpoint and model with per-input pricing and batch question handling.
Trend signal
This release supports model-answers-become-task-interfaces: Google is presenting model output as bounded scores and choices for defined tasks, rather than open-ended generated text. The signal is concrete at the API level, but the announcement alone does not establish how widely teams will adopt such interfaces or whether they outperform simpler classifiers in production.
Why now
Google frames the API around operational latency and cost: multiple checks can share one multimodal input, and confidence values can route uncertain results to a fallback or reviewer. The announcement’s examples show Google applying the endpoint inside existing products, while developer access and SDK integrations extend that decision pattern outward.
Why it matters
A bounded decision endpoint can simplify systems that currently call a generative model repeatedly to classify, route, or grade the same material. If the shared input is genuinely billed once and question-level overhead is small, teams may be able to consolidate checks and make cost more predictable. Confidence scores also provide a natural handoff mechanism: automate clear cases, reserve human review for ambiguous ones.
The trade-off is that a specialized endpoint becomes another dependency in the model stack, and an API confidence score is not itself proof that a decision is correct or well calibrated for a team’s data. Builders should compare it with their current classifier or model call on representative inputs, including difficult and adversarial cases, and measure end-to-end outcomes rather than assuming Google’s benchmark ranking transfers to their workload.
Limits and uncertainty
The latency, calibration, benchmark leadership, and nearly fivefold cost comparison are claims in Google’s announcement; the announcement does not independently establish their performance on a particular customer workload. It also does not report production error rates for the named internal uses. The confidence score’s relationship to real-world correctness will depend on task and data, and teams still need to determine thresholds and review policies. The page describes ecosystem integrations as rolling out and Google Cloud support as coming soon, so their availability may vary.
Skill to build now
Take one repetitive classification or routing task and express it as a bounded decision: define the input, candidate choices or rubric, and the cases that must go to a human. Run the same examples through your current system and a Decisions API prototype, then compare accuracy, latency, and total cost—including review and fallback work—before replacing anything.
Reader takeaway
Gemini Decisions API turns multimodal classification and scoring into a dedicated, batchable API primitive with published per-input pricing. Treat it as a candidate control-plane component, not a proven drop-in: test the claimed speed, calibration, and savings against your own tasks before routing production decisions through it.