Google Releases Gemini Decisions API: A New Tool for Multimodal Probabilistic Decision-Making

Google launches the Gemini Decisions API, supporting multimodal batch closed-ended decision-making with calibrated probabilities.
Google released the Gemini Decisions API, powered by the Gemini Decision 1 model, which transforms large model reasoning into quantifiable decision signals. Instead of outputting free text, the API returns calibrated probabilities for user-defined closed-ended questions, supporting up to 1,000 yes/no, multiple-choice, or rating-scale questions per request, with inputs covering text, images, audio, video, and PDFs. Priced at $0.10 per million input tokens—at the lower end of Google's product line—it is suitable for large-scale batch decision scenarios such as content moderation, document classification, and multimedia screening. The article notes that calibrated probabilities facilitate direct threshold setting and risk weighting downstream, but the model's reliability in specialized fields lacks benchmark data support, and closed-ended questions must be manually predefined, making it unsuitable for open-ended tasks.
Title
Google Releases Gemini Decisions API: A New Tool for Multimodal Probabilistic Decision-Making
Meta Description
Google launches the Gemini Decisions API, powered by Gemini Decision 1, supporting text, image, audio, video, and PDF inputs, returning calibrated probabilities, handling up to 1,000 yes/no, choice, or rating questions per request, priced at $0.10 per million input tokens.
Article Content
Google recently launched the Gemini Decisions API, powered by the Gemini Decision 1 model, which returns calibrated probability results for various inputs including text, images, audio, video, and PDFs, supporting up to 1,000 yes/no, choice, or rating-scale questions. Pricing is $0.10 per million input tokens.

Core Capabilities of the Multimodal Decision API
The Gemini Decisions API is designed to transform the reasoning capabilities of large models into quantifiable decision signals. Unlike conventional text generation interfaces, it does not output free-form answers; instead, it provides probability judgments for user-defined closed-ended questions. These questions can be simple yes/no judgments, multiple-choice selections, or scoring according to a rating scale.
A single request supports up to 1,000 questions, meaning developers can perform batch decision evaluations in one call. Input modalities cover text, images, audio, video, and PDFs, enabling the API to be applied to scenarios such as content moderation, document classification, and multimedia asset screening. The returned "calibrated probabilities" are a notable design choice: calibration means the model's output probability values are statistically closer to true frequencies, making it easier for downstream systems to directly set thresholds or perform risk weighting.
Calibrated Probability refers to a model's output probability values being statistically consistent with real-world event frequencies. For example, if a model assigns a probability of "0.8" to 100 different questions, then approximately 80 of those 100 questions should have a true answer of "yes." Uncalibrated models often exhibit overconfidence or underconfidence—for instance, outputting 0.95 for a judgment where the actual confidence is only 60%. Calibration is typically achieved through post-processing techniques such as temperature scaling or isotonic regression, or through specialized training on large-scale annotated data. For downstream systems, well-calibrated probabilities can be directly used to set thresholds, calculate expected losses, or perform risk weighting without requiring an additional mapping from scores to probabilities. In scenarios like content moderation, this means operations teams can directly determine cutoff points based on the ratio of "cost of false negatives" to "cost of false positives."
Pricing Strategy and Cost Considerations
At $0.10 per million input tokens, the pricing sits at the lower end of Google's model product line. For tasks requiring large-scale, high-frequency binary judgments, this pricing structure may offer a cost advantage over calling a general-purpose large model and then parsing the output. However, it should be noted that the materials only disclose input token pricing, without mentioning output billing or whether other additional fees apply. When evaluating total costs, developers still need to calculate based on the number of questions and the length of input content.
From a product positioning perspective, this pricing appears more like "infrastructure-type" pricing for batch decision scenarios rather than pricing for conversational interaction. For teams already using Google Cloud or the Gemini ecosystem, the integration barrier is relatively low.
The billing difference between input tokens and output tokens is a common structure in LLM API pricing. Tokens are the smallest units the model processes for text, roughly equivalent to a word or part of a word; multimodal inputs (images, audio, video) are typically encoded into a certain number of tokens. Most generative APIs price input and output separately, with output unit prices often significantly higher than input, because autoregressive generation requires per-token computation. The Gemini Decisions API's pricing is based only on input tokens, which is consistent with its output format—it returns probability values for each question, with extremely small and structurally fixed output volume, so it may not be billed or may incur only minimal fixed overhead. When estimating costs, the main variables for developers are the total token volume of input content per request multiplied by the processing overhead corresponding to the number of questions, rather than the unpredictable output length in traditional generation tasks.
Use Cases and Potential Limitations
Typical uses of a probabilistic decision API include: compliance judgments on user-submitted content, prioritizing large volumes of documents according to rating criteria, and multi-label screening of multimedia assets. Because it returns probabilities rather than definitive answers, systems can set conservative or aggressive thresholds to flexibly balance recall and precision.
Limitations also exist. The actual quality of calibrated probabilities depends on the model's performance on domain-specific data, and the materials do not provide benchmark data, so its reliability in specialized fields cannot be assessed. Additionally, closed-ended questions must be manually predefined, making it unsuitable for open-ended, exploratory tasks. The specific format requirements and size limits for multimodal inputs are also not specified in the materials.
The trade-off between recall and precision is a core framework for evaluating binary decision systems. Recall refers to the proportion of all true positives that are correctly identified, while precision refers to the proportion of samples judged as positive that are actually positive. The two typically trade off against each other: lowering the decision threshold can capture more positives (higher recall) but introduces more false positives (lower precision). The value of a probabilistic API lies in giving business teams explicit control over this trade-off—for example, in content moderation for violations, if the cost of missed detections is high, lower the threshold to preserve recall; if the cost of false positives is high, raise the threshold to preserve precision. Because it returns continuous probabilities rather than hard labels, the same interface can be configured with different thresholds across business lines without retraining the model.
Implications for Developers
The Gemini Decisions API represents a trend: encapsulating large model capabilities into narrower, more controllable interfaces to reduce integration complexity. Developers no longer need to handle parsing and error tolerance for natural language output; they directly obtain structured probability values. For teams looking to embed AI judgments into business processes, this type of API can shorten the path from prototype to production. Whether to choose this API depends on how well the specific task matches the calibration quality, modality coverage, and cost structure.
One-line Summary
Google launches the Gemini Decisions API, supporting multimodal batch closed-ended decision-making with calibrated probabilities.
Paragraph Summary
Google released the Gemini Decisions API, powered by the Gemini Decision 1 model, which transforms large model reasoning into quantifiable decision signals. Instead of outputting free text, the API returns calibrated probabilities for user-defined closed-ended questions, supporting up to 1,000 yes/no, multiple-choice, or rating-scale questions per request, with inputs covering text, images, audio, video, and PDFs. Priced at $0.10 per million input tokens—at the lower end of Google's product line—it is suitable for large-scale batch decision scenarios such as content moderation, document classification, and multimedia screening. The article notes that calibrated probabilities facilitate direct threshold setting and risk weighting downstream, but the model's reliability in specialized fields lacks benchmark data support, and closed-ended questions must be manually predefined, making it unsuitable for open-ended tasks.
Related articles

AI Could Overthrow Governments in Six Months? ARC Head Issues His Starkest Warning Yet
ARC head Paul Christiano warns AI capable of overthrowing governments may arrive in 6-12 months, with ~one-third odds of superintelligence within a year. He urges frontier labs to pause algorithmic efficiency gains and fix safety first, and reveals how the Hugging Face incident foreshadowed AI taking over evaluation.

How Do Spies Infiltrate Nvidia? A CIA Agent Reveals a Honeypot Operation
A former CIA agent reveals in a YouTube interview that intelligence agencies ran a honeypot seduction operation against Nvidia CEO Jensen Huang, exposing the hidden spy risks facing tech executives.

How NVIDIA Cosmos Helps Delivery Robots Rehearse Unexpected Situations in Advance
NVIDIA Cosmos World Foundation Model expands a small amount of real input into large-scale synthetic scenarios, enabling delivery robots to train for unexpected situations in simulation in advance. Cocoa Robotics uses it to cover unexpected events and combines Isaac Sim and Isaac Lab to evaluate perception, prediction, and path planning.