Google joins the decision model battle: supports five modalities, costs more than twice as much as Jev

Beating AI News Flash: Google releases its first decision model, Gemini Decision 1, supporting mixed input of text, images, audio, video, and PDF, with a 256,000 Token context.

Compared with Jev and the OpenAI Decisions API, Google can directly process more types of materials. For example, a customer service recording, order information, and product photos can be submitted to the model together to simultaneously determine the customer's request, the severity of the issue, and whether human intervention is needed.

It can judge up to 1,000 questions at a time, with multiple questions sharing the same material and requiring payment for the material input fee only once. The officially announced end-to-end response time is 150 to 250 milliseconds. Within 16,000 Tokens, when the number of questions increases from 1 to 100, the model's computation time remains basically unchanged.

In Google's internal testing, Gemini Decision 1 scored 92.2% on JevBench, surpassing Jev 1.14's 90.0% and the OpenAI Decisions API's 90.9%. However, on the tool-calling benchmark When2Call, Google lagged behind Jev's 82.0% with 79.9%.

In third-party Artificial Analysis's Decision Index v2.1, Gemini Decision 1 received an overall score of 65.98, ranking ninth among 18 models, roughly on par with Jev's 66.23 and behind the leader Drex Max v1.1's 72.62.

Google charges $0.1 per million input Tokens, with output free. Calculated according to Artificial Analysis's unified task suite, for every 1 million tasks completed, Google costs about $345, Jev only $156, and Drex Max about $103. Google averages 211 milliseconds per task, while Jev takes 92 milliseconds.

Google has already used the new model for model routing in Gemini Notebook, reducing latency for some daily requests by up to 40%. The API is now open, with free testing quota available.