Google DeepMind launches Gemini Decision 1 for agent-based decision-making

Google DeepMind has launched Gemini Decision 1 and a dedicated Decisions API, aiming to extract numerous closed-loop judgments from general generative models—such as “which tool to call,” “whether it complies with policies,” “which model category to assign,” and “how high the risk is.” Developers can submit up to 256,000 tokens of text, images, audio, video, or PDFs in a single request and pose up to 1,000 independent questions about the same material. The model does not generate word-by-word explanations but instead provides probabilities and confidence scores for options in a single forward pass. Google reports typical end-to-end latencies of 150 to 250 milliseconds, with pricing at $0.10 per million input tokens and free output. However, these latency figures are primarily based on tests involving no more than 16,000 tokens and up to 100 questions, and do not guarantee the same speed under full context with 1,000 questions. Independent evaluations ranked it 9th among 18 decision models, indicating strong calibration and multimodal capabilities, but not overall superiority across all decision tasks.

Article author, source: Google DeepMind

Many agents today invoke full generative models even when they only need to answer "yes" or "no." The model must read the context, perform reasoning, and output text token by token, after which the application must parse a boolean value or category from the response.

This approach is flexible but inefficient in high-frequency workflows. A customer service agent may need to sequentially assess user intent, account risk, applicable policies, tool permissions, escalation queues, and response models; calling a generative model at each step accumulates latency, costs, and formatting errors.

Gemini Decision 1 takes a different approach: developers first define a limited set of candidate answers, and the model is solely responsible for scoring these options. It is not a chat model, does not draft explanations, and functions more like a general-purpose layer for classification, ranking, and strategic judgment, capable of understanding long contexts and multimodal information.

Google offers three question formats:

• Yes/No judgment, returning the probability that the condition is met;
• Select from 2 to 255 options and return the normalized probability for each option;
• Score according to a predefined ordered scale, while returning the probability of each level and the weighted score.

Each result also includes a confidence score between 0 and 1. The application can set a threshold: automatically execute when confidence is high, or route to a stronger model or human review when confidence is insufficient.

A single request can contain up to 256,000 tokens and pose 1,000 independent questions on the same input.

For example, an insurance company can submit a scanned claim form, on-site damage photos, a recording of the customer call, and account information all at once, and request the model to determine whether the materials are complete, classify the type of incident, detect any signs of fraud, assign it to the appropriate queue, and prioritize its urgency.

Shared materials are charged only once, and each question adds only its own instructions and options in tokens. Therefore, when dozens of checks need to read the same long file, it is more cost-effective than sending the file dozens of times. The official pricing is $0.10 per million input tokens, with no charges for output tokens, and no separate fees for OCR or audio transcription.

But these 1000 questions must be independent of each other. The model will not allow the second judgment to read the result of the first judgment. If the workflow is “first determine the document type, then apply different review rules based on the type,” developers still need to connect the two steps in code or call the API again. When generating explanatory text, extracting arbitrary fields, or drafting responses, generative or information extraction models are still required.

Gemini Decision 1 skips token-by-token decoding and scores all candidates in a single forward pass. Google reports that the median model computation time is 50 to 100 milliseconds, with end-to-end latency, including service and network overhead, approximately 150 to 250 milliseconds.

Official testing shows that, with multimodal inputs of up to 16,000 tokens, increasing the number of questions per request from one to 100 results in nearly flat latency. In the same benchmark, the cost is reportedly nearly five times lower than calling Gemini 3.5 Flash-Lite via the Interactions API.

These figures cannot be directly extrapolated to all scenarios. Although the model supports 256,000 tokens and 1,000 questions, Google has not demonstrated that simultaneously reaching both limits can still be completed within 250 milliseconds; upload times for large videos, audio files, and PDFs are not included in the model’s computation time. Actual deployment must re-measure using your own file lengths, number of questions, and network locations.

Traditional classifiers are typically designed for a single, pre-defined data format, while Gemini Decision 1 can simultaneously process text, images, audio, video, and PDFs within the same request.

The official example submitted to the model a photo of a kitchen counter and a text description, asking three questions: whether there is any uncleaned liquid, whether to wipe it down or call cleaning staff, and the severity of the situation. The model determined there was a small spill, recommended wiping it with a cloth, and provided probabilities for each option.

Developers can also reduce the resolution of images or videos to decrease token usage and costs at the expense of some detail. For claims processing, content moderation, product inspection, or video event recognition, this approach is easier to implement than running OCR, speech recognition, vision models, and rule engines separately; however, a unified model also means that if it misinterprets visual or audio content, multiple subsequent decisions may be affected simultaneously.

Google states that Gemini Decision 1 leads in routing, reasoning, ranking, and multimodal understanding benchmarks, achieving the lowest calibration error. In its official comparisons, the model scores 92.2% on JevBench, surpassing GPT-6 Luna’s 90.9% and Jev 1.14’s 90.0%; its calibration error is 0.0250, approximately half of Jev’s 0.0494.

But it doesn't dominate all tasks. Jev scores slightly higher than Gemini at 86.8% versus 86.2% on the browser task decision benchmark Mind2Web Multimodal, and leads at 82.0% versus 79.9% on When2Call, which evaluates when to invoke tools.

The standalone Artificial Analysis Decision Index v2.1 scores Gemini Decision 1 at 66 out of 100, ranking it 9th among 18 decision models. It leads in knowledge and reasoning tasks, but in the AA-Paperclip test—which requires correctly understanding entire resumes, invoices, receipts, or papers—only 33% of documents were answered completely correctly, compared to 63% for Drex Max v1.1. It also trails behind Drex Max and Jev in the standalone When2Call test.

This shows that “well-calibrated confidence” and “each judgment being most accurate” are not the same thing. A well-calibrated model is more aware of when it might be wrong, which helps set thresholds for human review; however, if the application incorrectly treats high confidence as a guarantee of correctness, it may still automatically make erroneous decisions.

Google has integrated this API into multiple products. Gemini Notebook uses it to determine which model and level of reasoning to apply for a query, with the company reporting up to a 40% reduction in latency for general queries; Stitch uses it to quickly respond to design changes triggered by voice or clicks; Google Labs' CC selects reactions from over 200 emojis; and Google Search places it before MCP tool calls to detect adversarial behavior and harmful content.

These cases reveal its true positioning: it is not the intelligence users ultimately see, but the control plane behind the Agent. Generative models are responsible for writing code, organizing language, and solving open-ended problems; decision models are responsible for determining what to call next, whether to allow a call, which model strength is needed, and when to stop automation and hand over to a human.

This division of labor may be more suitable for large-scale Agent systems than simply scaling up general-purpose models, as many routine steps do not require “better writing,” only fast, cost-effective, and probabilistically reliable branch selection.

Gemini Decision 1 is currently available via the Gemini Developer API's `v1beta` endpoint, with Google Cloud support still labeled as "coming soon." Google has not publicly disclosed parameters, training data, training compute, specific architecture, or full system cards, making it impossible to determine whether it is distilled from existing Gemini models or to independently verify safety and data claims at the training level.