Gemini Decision 1: Features, Benchmarks, Pricing, and Availability
Google introduced Gemini Decision 1 on November 11, 2026, entering the growing market for AI decision models. Unlike traditional large language models that generate text, Gemini Decision 1 evaluates options and returns structured answers with confidence scores.
It supports text, images, audio, video, and PDFs, making it useful for automated decisions, policy checks, content classification, and workflow routing.
With a reported input price of just $0.10 per million tokens and no output-token charges, Gemini Decision 1 offers a different approach to building AI-powered applications. In this article, we'll explore Gemini Decision 1's key features, multimodal capabilities, benchmark performance, pricing, availability, and how it compares with other decision models such as Jev and GPT-6 Luna.
What You Need to Know
|
What Is Gemini Decision 1?
Gemini Decision 1 (gemini-decision-1) is Google's first specialized decision model, designed to evaluate structured questions and return probabilities rather than generate text.
It is available through Google's new Decisions API, which is separate from the Interactions API used for generative Gemini models.
The difference is straightforward. A traditional AI model might generate a response such as "billing" and require an application to interpret that text. Gemini Decision 1 evaluates the available choices directly and returns a probability for each option.
This eliminates the need to parse generated text and avoids output-token costs.
One of its key strengths is calibration. Google's evaluations show that Gemini Decision 1 has the lowest Expected Calibration Error (ECE) among the compared models. This means its confidence estimates more closely reflect how often its predictions are correct.
Gemini Decision 1 Key Features
Gemini Decision 1 can evaluate between 1 and 1,000 questions against a shared input in a single request, subject to its 262,144-token context window. It supports three question types:
Each question is evaluated independently against the shared input. This approach can reduce processing costs when an application needs to check one document against multiple rules. For example, a customer support message could be evaluated against 40 different policy requirements in one request. Google also reports that latency remains nearly flat when increasing the number of questions from 1 to 100. |
One of Gemini Decision 1's biggest advantages is its multimodal capability. It can evaluate different media formats directly without requiring separate transcription or optical character recognition (OCR) steps. Possible applications include:
Google says there are no separate audio transcription or OCR charges. Image and video resolution can also be adjusted to balance visual detail and token consumption. There are some limitations: question instructions must be text-based, and YouTube URLs and direct Google Cloud Storage (gs://) URIs are not accepted. Media must be provided through the Files API, HTTPS URLs, or inline data. |
Every answer includes a confidence score between 0 and 1, representing the estimated probability that the selected answer is correct. Applications can use these scores to decide what happens next:
Google recommends setting confidence thresholds using labeled examples from the application's own data. A model's calibration on published benchmarks does not guarantee identical performance on every business dataset. |
Gemini Decision 1 can also act as a decision layer in front of generative AI models. Instead of sending every request to an expensive reasoning model, an application can first classify the request and determine which model or processing level is appropriate. Google reports that Gemini Notebook uses this approach to route queries to different model tiers and thinking levels, reducing latency for everyday queries by up to 40%. Google Search also uses the Decisions API as a safety layer for Model Context Protocol (MCP) tools to identify potentially adversarial behavior. This makes decision models useful wherever an application needs classification, routing, or rule-based evaluation rather than a generated response. |
Gemini Decision 1 Benchmark Performance
Google reports that Gemini Decision 1 leads in 7 of its 9 accuracy benchmarks and has the lowest calibration error among the compared models.
The comparison includes TypeSafe Jev 1.14, GPT-6 Luna through OpenAI's Decisions API, and Gemini 3.5 Flash-Lite.
Routing, ranking, and reasoning benchmarks
Benchmark | What it measures | Gemini Decision 1 | Jev 1.14 | GPT-6 Luna | Gemini 3.5 Flash-Lite |
JevBench | Decision-making performance | 92.2% | 90.0% | 90.9% | 86.6% |
Amazon ESCI | Product search relevance | 58.7% | 56.7% | 54.7% | 47.2% |
New Yorker Caption | Caption matching | 79.7% | 72.0% | 68.4% | 73.7% |
Mind2Web Multimodal | Selecting the next web action | 86.2% | 86.8% | 81.8% | 78.0% |
PubMedQA | Biomedical research questions | 79.8% | 78.2% | 74.4% | 77.6% |
Global-MMLU-Lite | Multilingual knowledge | 86.0% | 85.0% | 79.6% | 78.8% |
When2Call | Tool-call decisions | 79.9% | 82.0% | 70.4% | 60.9% |
Gemini Decision 1 performs strongly across most of these evaluations. However, Jev 1.14 leads on Mind2Web Multimodal and When2Call.
The When2Call result is particularly relevant for agent-based applications because it evaluates whether a model should call a tool and which action it should take.
Audio and video performance
Gemini Decision 1 also demonstrates multimodal capabilities in Google's evaluations.
Benchmark | What it measures | Gemini Decision 1 | Gemini 3.5 Flash-Lite |
MMAU | Audio understanding | 71.8% | 67.7% |
Video-MME | Video understanding without subtitles | 69.1% | 63.3% |
Jev and GPT-6 Luna do not have reported scores on these media benchmarks in Google's comparison.
These results highlight one of Gemini Decision 1's main advantages: evaluating audio and video directly rather than relying only on text-based inputs.
Calibration and cost comparison
Calibration measures how closely a model's confidence scores match its actual accuracy. Lower ECE values indicate better calibration.
Metric | Gemini Decision 1 | Jev 1.14 | GPT-6 Luna | Gemini 3.5 Flash-Lite |
Top-1 ECE (lower is better) | 0.0250 | 0.0494 | 0.0721 | 0.2104 |
Total price across 7 shared evaluations | $0.88 | $0.68 | $0.93 | $4.14 |
Gemini Decision 1 has approximately half the calibration error of Jev 1.14.
However, Jev is the cheapest option across the seven shared evaluations. Gemini Decision 1 costs approximately 30% more than Jev in this comparison, but offers better calibration and native media support.
Google's claim that Gemini Decision 1 is nearly five times cheaper refers to its comparison with Gemini 3.5 Flash-Lite, not Jev.
Gemini Decision 1 Pricing and Availability
Gemini Decision 1 has a straightforward pricing structure because it does not generate output tokens.
Pricing component | Cost |
Input per 1M tokens | $0.10 |
Output tokens | $0 |
Audio, video, image, and PDF input | Same input rate |
The model is available through the Gemini Developer API using the v1beta version.
Free and paid tiers
Feature | Availability |
Free tier | Google AI Studio |
Free-tier rate limit | 200 requests per minute |
Free-tier daily limit | 10,000 requests |
Paid Tier 1 | 1,500 requests per minute |
Paid Tier 3 | Up to 15,000 requests per minute |
There are some limitations:
Context caching is not supported.
The Batch API is not supported.
Flex and Priority inference are not supported.
Repeated inputs across separate requests are billed again.
The lack of context caching is important for applications that repeatedly evaluate the same long document. Combining multiple questions into one request can help avoid paying repeatedly for the same input.
Google also distinguishes between free and paid-tier data usage: free-tier data may be used to improve Google's products, while paid-tier data is not.
Conclusion
Gemini Decision 1 introduces a specialized approach to AI-powered decision-making. Instead of generating text, it evaluates structured questions and returns probabilities that applications can use for routing, classification, policy checks, and automated workflows.
Its strongest advantages are native multimodal support, the ability to evaluate up to 1,000 questions in one request, and its reported calibration performance.
However, it is not a clear winner in every category. Jev 1.14 remains cheaper across Google's shared evaluations and performs better on tool-call decisions.
For developers building text-based agent routers, Jev remains a competitive option. For applications involving images, audio, video, PDFs, or confidence-based decisions, Gemini Decision 1 is worth evaluating.
Comments