Amazon Launches the Nova Model Family at re:Invent

At re:Invent, Amazon entered the foundation model race with a family of its own, Nova, delivered through AWS Bedrock. It is more than the three text models it usually gets reduced to. Nova Micro is text-only, fast and cheap, with a 128,000-token context window, aimed at high-volume work like classification and extraction. Nova Lite is the low-cost multimodal option: it reads images and documents as well as text, with a 300,000-token context. Nova Pro is the capable tier, also multimodal at 300,000 tokens, and Amazon states it is equal to or better than GPT-4o on 17 of 20 benchmarks. A fourth and larger model, Nova Premier, was announced at the same event but scheduled for the first quarter of 2025 rather than shipping alongside the others. Two generative models came with the launch as well: Nova Canvas for images and Nova Reel for video. Nova launched on Bedrock, and in practice that was how you used it, though Amazon's announcement stops short of calling it exclusive. For companies already running on AWS the appeal is placement: the models sit next to S3, Lambda and the rest, so data does not have to leave the account to reach a capable model. The area where Nova Pro's advantage is clearest is document understanding, which is also the least glamorous and most common enterprise workload.