Newshot Gemini Decisions API: no answers, just multimodal decisions
Google has launched the Gemini Decisions API, which uses Gemini Decision 1 to evaluate up to 1,000 closed-ended questions at once across text, images, audio, video and PDFs, returning probabilities and confidence scores. It focuses on low latency, low calibration error and zero output-token billing, making it suitable for tool routing, rule checks, content scoring and triaging items for human review.
- Evaluates up to 1,000 questions in a single request
- End-to-end latency of 150 to 250 milliseconds
- Priced at $0.10 per million input tokens
A call recording, a photo, a scanned PDF — the program has to decide which tool to hand it to, whether to escalate to a human, and what score to assign. Google's new Gemini Decisions API does not write answers: a single request can evaluate up to 1,000 closed-ended questions and directly returns probabilities and confidence scores. Judgments are pulled out of text and turned into numbers code can read.
Reads five kinds of material at once, computes 1,000 judgments together
The Gemini Decisions API does not generate a passage of answer text; instead, it assigns a probability to each candidate within a single request, aimed at moments when a system needs to "pick one" or determine "whether it complies."
Google DeepMind released the Gemini Decisions API on November 11, 2026. The underlying model is called Gemini Decision 1, with the model identifier gemini-decision-1 (the model name callable in the Gemini API). It is available through the Gemini API and Google AI Studio, with initial uses including tool selection, policy checks, model routing and content scoring. "Routing" here (sending requests to different models based on the task) does not have the model write an answer; instead, the program uses the returned scores to decide which model to call next, which tool to choose, or whether to let something through.
A single request can contain text, images, audio, video and PDFs, with a shared input limit of 256K tokens (the unit of measure models use when processing text). Resolution for images and video can also be adjusted, trading visual detail for input cost. The same material can be paired with up to 1,000 independent questions, which can be yes/no, multiple-choice or scale ratings, and every candidate result receives a probability and a confidence score.
Text, images, audio, video and PDFs share a single input.
Candidates are returned one by one with probabilities and confidence scores.
Total input is measured in tokens.
The most obvious difference from ordinary generative interfaces is that it does not generate an answer token by token. Gemini Decision 1 completes candidate scoring in a single forward pass (the model performs one round of computation over all the input), then hands the scores to the program to execute branches. The official end-to-end latency is 150–250 milliseconds, with a median model computation time of 50–100 milliseconds. Latency barely grows with the number of questions — that is, within the official test range of 1–100 questions; whether it stays flat at 1,000 questions is not stated in the materials.
This interface combines "understanding the material" and "making the choice" into a single request. After developers submit content and questions, what they get back is not a natural-language answer but a set of executable choice signals. Whether judgment quality changes with the scale of questions — the official materials provide no comparative data at 1,000 questions.
The answer is not a passage of text, but a number that can go straight into code
The same relevance described above also applies to closed-ended judgments over images, audio, video, PDFs and text.
Developers hand the same material to Gemini Decision 1 and ask multiple independent questions at once, including true/false judgments, multiple choice and scale ratings. Each item returns a confidence score from 0 to 1, and the program uses it to choose a path: continue automatic processing, hand off to another set of tools, pass to a heavier model, or send to human review. The confidence score here is the model's judgment score for an option, not a guarantee of "answer accuracy."
This kind of interface targets the "fork in the road" inside AI agents. When a system needs to determine whether an image violates moderation rules, which classifier a recording is better routed to, or what level a piece of content falls into on a rating scale, none of it requires the model to generate an answer word by word. Probabilities can go directly into code, eliminating the layer where developers parse natural language themselves and then convert the answer into branch conditions.
Less output actually makes it better suited to automated systems. Gemini Decision 1 does not decode an answer token by token; it scores candidates in a single forward pass. That removes a layer of intermediate text meant for humans to read but easy to misread — what automated systems want has never been an explanation, but a number they can branch on.
Tests cover inputs of no more than 16K tokens, and latency stays essentially flat as questions increase from 1 to 100. Billing also centers on input: output tokens are not billed, and a request is charged by input regardless of how many questions are evaluated, with no separate speech transcription or OCR charges.
The official comparison is that on a shared benchmark suite it is nearly five times cheaper than Gemini 3.5 Flash-Lite paired with the Interactions API; this ratio comes from Google's own tests and would need to be recalculated for very short inputs with very few questions.
For ordinary users, the impact will be hidden in the background: image moderation gets triaged earlier, voice requests get routed faster, and content scoring no longer rests on a passage of text that cannot be reliably parsed. But a confidence score is not the fact itself; a low score does not automatically mean a violation, and a high score does not mean human fallback can be removed.
The remaining ~7 minutes are for members.
The free portion ends here. Behind the paywall there is also:
Paywalled highlight: AI no longer writes answers — are models starting to make choices for programs?