to leave a comment.
![[AI Office] Google 'Gemini Decisions', 1,000 judgments in around 0.2 seconds... 5 times cheaper than lightweight models](/_next/image?url=http%3A%2F%2Fwww.coinreaders.com%2Fdata%2Fcoinreaders_com%2Fmainimages%2F202611%2F2026111702341278.png&w=3840&q=75)
Google DeepMind announced the official launch of 'Gemini Decisions API', which instantly discriminates input data. This method provides only answers instead of generating sentences, and the dedicated model that drives it is 'Gemini Decision 1'. A path has been opened for companies that have entrusted even tasks requiring only 'yes or no' decisions, such as spam email classification or receipt rule verification, to chatbots, to reduce costs.
Until now, generative AI has long elaborated sentences word by word even for simple judgments, such as determining 'suitable' or 'unsuitable'. As a result, output token fees were incurred, and responses were delayed by several seconds.
According to Google DeepMind, the new API scans text, image, audio, video, and PDF data in a single forward pass. It can simultaneously determine up to 1,000 yes/no and multiple-choice questions and can also handle scoring according to rubrics.
The key is speed and price. The model does not write text but outputs the confidence level for each option as a probability between 0 and 1. The time it takes to process from beginning to end is only 150-250 milliseconds (0.15-0.25 seconds).
The fee is $0.10 (approximately 140 KRW) per 1 million input tokens. Output tokens are not charged. Google explained that it is about 5 times cheaper than the latest lightweight models in the Gemini Flash series.
This model will be used for tasks that require 'decisions' rather than 'words'. Representative examples include large-scale content review, financial transaction anomaly detection, and internal network security filtering. It also includes filters that real-time screen malicious sites or phishing links during web searches, and automatic classification of tens of thousands of consultation tickets coming into customer service centers.
However, it is not an all-purpose tool. It is not intended for writing proposals or having emotional conversations, but rather specialized for closed questions with pre-defined options. A 'hybrid' structure is realistic, where existing inference models handle report drafts, and the decision model handles only suitability verification and initial screening of large data.
Practitioners don't have much to do right away. They can first check whether simple suitable/unsuitable judgments can be detached from generative chatbots in internal tasks. Then, they can confirm pricing policies and rubric functions in Google Cloud and the API console, and try applying the ultra-low-cost model to automatic customer inquiry classification or spam blocking filters.
Domestic startups and corporate practitioners facing high generative AI operating costs may consider designing for specific purposes, such as entrusting conversations to general chatbots and large-scale classification and judgment to decision-specific models.
Newsletter
Get key news delivered to your email every morning
to leave a comment.