At the re:Invent2025 conference, Amazon Web Services (AWS) launched four "Nova2" series self-developed large models, covering multi-modal scenarios such as text, images, videos, and speech. For the first time, they have built-in web search and code execution capabilities, claiming to achieve "industry-leading price-performance ratio" for the same tasks.
Performance Comparison
- Nova2Lite: Positioned as a high-cost-performance inference model, it matches or exceeds Claude Haiku4.5 in 13 out of 15 benchmarks, and matches or exceeds GPT-5Mini in 11 out of 17 benchmarks, with a cost of about 50% of the latter.
- Nova2Pro: Designed for complex Agent tasks, it matches or exceeds Claude Sonnet4.5 in 10 out of 16 evaluations, and matches or exceeds Gemini3Pro Preview in 8 out of 18 evaluations.
- Nova2Sonic: An end-to-end speech model, with real-time latency below 600ms, supporting a context of up to one million tokens and asynchronous background tasks.
- Nova2Omni: The industry's first unified multi-modal model, capable of inputting text/images/videos/audio and outputting text + images, completing understanding and generation with a single model.
Technical Highlights
All models in the series are integrated with "web search + code execution" dual tools, enabling real-time internet information retrieval and Python execution, ensuring answers based on the latest facts rather than just training data. AWS stated that tens of thousands of enterprises have already used the Nova series for content production, multi-step automation, and AI Agent development.
Market Strategy
AWS also launched the "Nova Forge" custom training service, which costs $100,000 annually to inject private data during the pre-training or post-training phase, building a tailored cutting-edge model. The goal is to reduce the cost of enterprises building large models from "hundreds of millions of dollars" to the million-dollar level.
Industry Perspectives
