Introduction

In the November 11, 2026 release notes, the Gemini API added the decision-task-specific model “Gemini Decision” (gemini-decision-1) and the new Decisions API.

Until now, decisions and classifications—such as “Does this image violate the rules?”, “Which category does this inquiry fall into?”, or “Does this response meet the criteria?”—have generally been made by prompting a generative LLM and having it return an answer in text or JSON. Gemini Decision specializes in these use cases and has the following three characteristics:

• High speed
• Calibrated (adjusted so that confidence levels match actual accuracy rates)
• Zero output tokens

📌 Who is affected

• Developers operating content moderation, inquiry routing, and quality evaluation (LLM-as-a-Judge) using LLMs
• Teams concerned about output-token charges and latency for each decision
• People building multimodal decision pipelines that include images, audio, video, and PDFs

Note that the deprecation of existing features and breaking changes are not included in the data for this release, and this is not a change that requires action (action_required: false). However, it is a release worth evaluating and comparing as an option.

⚠️ Premises of this article
The specific fee amounts and rate-limit values, as well as the request/response specifications of the Decisions API, are not included in the input data for this article. Therefore, this article will not include code examples and will clearly identify the points that should be checked in the official documentation.

Overview of the changes

Differences between the conventional approach and the Decisions API

Supported decision formats

(Each “example” is a conceptual illustration of a use case and is not an official example.)

Changes

What was added

Contents

Item
New model	Gemini Decision (Model ID: gemini-decision-1)
New API	Decisions API
Decision format	Binary / Categorical / Rubric
Input	Text, images, audio, video, and PDFs
Output characteristics	Calibrated results, zero output tokens
Provider	Gemini API

Breaking down the three features

1. Zero output tokens
Generative LLMs generate “decision rationales” or “JSON” as text, so the corresponding output tokens are subject to charges and latency. Gemini Decision is stated to have zero output tokens, and compared with conventional approaches, it is expected that there will be no charges for output tokens and that latency will also be lower (this is an “expectation” based on the data; please verify it through your own testing).

2. Calibrated confidence
A longstanding challenge with generative LLMs is that even when they output a “confidence of 0.9,” their actual correctness rate is not necessarily 0.9. Calibration means adjusting confidence to align it with the actual correctness rate. For example, a group of decisions with a confidence of 0.9 is adjusted so that roughly 90% of them are correct. This makes it easier to design operational thresholds, such as routing only low-confidence cases to human review.

3. Multimodal input
In addition to text, images, audio, video, and PDFs can be evaluated directly as-is.

Pricing and rate limits

Dedicated sections are available.

• Pricing: /docs/pricing#gemini-decision
• Rate limits: /docs/rate-limits#gemini-decision
• Integration steps: Decisions API guide /docs/decisions-api

Specific pricing and limit values are not included in the diff data for this round. Be sure to check the official page before making an adoption decision.

Impact and Response

Who Should Consider This?

Use Cases	Points to Consider
Content Moderation (Including Images, Videos, and Audio)	Route cases between automated processing and human review using binary decisions and confidence thresholds
Automatic Classification of Inquiries and Documents	Category-based decisions may simplify output parsing
Quality Evaluation Using an LLM (LLM-as-a-Judge)	Rubric-based judgments may be well suited to scoring against evaluation criteria
Large-Scale Batch Decisions	Potential benefits of no output-token charges and low latency

Migration Decision Flow

Actions Developers Should Take

💡 Tips
No mandatory actions are required. We recommend starting by taking stock of your organization’s own decision tasks.

1. Decision Task Inventory: Identify all areas where your organization uses LLMs to classify, make decisions, or evaluate, and check whether they fit into the three types (binary/categorical/rubric-based).
2. Review the Official Documentation: Check the three items: the Decisions API guide, pricing, and rate limits.
3. Small-Scale PoC: Use existing decision data (with ground-truth labels) to compare accuracy, latency, and cost with the current approach.
4. Confidence Validation: Do not take “calibrated” at face value; use your organization’s own data to verify whether confidence levels correspond to actual accuracy rates.
5. Update the Knowledge Base: Add this to the team’s technology-trends memo as an option for decision, classification, and evaluation tasks.

Points to Note

• Because the output consists only of the decision result,if you need a natural-language explanation of “why that decision was made,” you may need to consider using a generative model in combination (this point should be confirmed in the official guide).
• Accuracy in multiple languages, including Japanese, input size limits, and similar details cannot be determined from the data available this time.

Summary

• On November 11, 2026, the Gemini API added a decision-specific model, gemini-decision-1, and the Decisions API.
• It supports three types of classification—binary, categorical, and rubric-based—and accepts text, images, audio, video, and PDFs.
• Its key features are high speed, calibrated outputs, and zero output tokens. It is expected to offer advantages in cost and latency over classification with conventional generative LLMs.
• There are no breaking changes or mandatory actions, but those with moderation, categorization, or evaluation pipelines may find a PoC worthwhile.
• Because pricing, rate limits, and API specification details are not included in the source data for this article, check the official documentation (/docs/decisions-api, /docs/pricing#gemini-decision, /docs/rate-limits#gemini-decision).

The direction of assigning judgment tasks to a dedicated decision model rather than “generation” is likely to become a new option for optimizing the cost of using LLMs.