Google has released a new Gemini Embedding — with multimodal embeddings!

Google has released a new Gemini Embedding — with multimodal embeddings! The new model can natively embed up to 120 seconds of video, 6 pages of PDF, and audio with text is also included. The model is available in both a free tier and a paid API. The embeddings, of course, are made in a nested manner — pieces of the embedding themselves are full-fledged embeddings, albeit less accurate. Google has again raised prices — for text they charge $0.2 per million tokens, while for multimodality the price somehow rises unhealthily — the price for video reaches up to $12 per million tokens (15k frames). Google is taking advantage of the lack of alternatives — other frontier companies are not engaged in this — the same OpenAI last updated embedding models in January 2024, also updating GPT 3.5 Turbo and GPT 4 Turbo.