AWS Updates Its Nova Models To Compete With Google, Anthropic and OpenAI

LAS VEGAS — At its re:Invent conference in Las Vegas today, AWS launched the latest generation of its family of Nova foundation models. Nova 2 Lite, Pro and Sonic — as well as a new Omni model that’s new to the family and that can take text, audio, images and video as inputs and output text and images. AWS announced the first set of Nova models at last year’s re:Invent and released a number of additional models for image generation, video generation and other use cases over time. For the most part, though, these models never quite achieved the mindshare of models from other vendors and by now, those original models have fallen well behind the state-of-the-art models from frontier labs like Anthropic, OpenAI and Google — as well as many open-weight models. With this new release, AWS is catching up, though, and the company isn’t afraid to cite benchmarks that put these models up against current frontier models (many of which only launched in the last few weeks). As for pricing, the new models will be competitively priced at $0.3/$2.5 per million input/output tokens for the Lite model. For the Pro and Omni models, AWS charges different prices depending on the format. For the Pro model, those start at $1.25/$10 per million input/output tokens for the text model. AWS’s Nova 2 model family (Credit: AWS). Nova 1 Lite The Lite version, for example, which AWS argues is ideally suited for use in customer service scenarios, document processing and workflow automation, is competitive and sometimes outperforms models like Anthropic’s Claude Haiku 4.5 and OpenAI’s GPT-5-mini (and Google’s Gemini Flash 2.5, but that’s an older model at this point). The core philosophy behind this model, similar to Anthropic’s Haiku family, is to find the right tradeoff between speed and cost. Nova 2 Lite benchmarks (credit: AWS). The Nova 2 Lite model also powers AWS’s new Nova Act service, its new service for managing browser-use agents. AWS says Nova Act delivers 90% reliability on the customer workflows it tested. Nova 2 Pro Nova Pro is AWS’s most intelligent reasoning model, the company argues, that excels at tasks that demand accuracy, multidocument analysis, reasoning over images and video, and coding tasks. The benchmarks put it in a league with Claude Sonnet 4.5 and GPT-5- and GPT-5.1, as well as Google’s Gemini 3 preview (with Gemini 3 only launching last week, AWS did not provide benchmarks against the final version of Gemini 3 yet). It would be nice to see a comparison with Anthropic’s Opus 4.5, since that beat Gemini 3 Pro on many benchmarks, but it’s worth noting that Nova 2 Pro only beat or matched Gemini 3 Pro on 8 out of 18 benchmarks. As AWS notes, November 2 Pro can also be used as a teacher model for knowledge distillation. Nova 2 Pro benchmarks (Credit: AWS). Nova 2 Sonic Sonic is a bit of a speciality model. It’s a speech-to-speech model that’s probably best suited for contact centre solutions like Amazon’s own Connect service and other AI assistant use cases. It’s meant for generating human-like conversations in real time and in multiple languages with expressive voices. It can switch between voice and text as needed and its token window is one million tokens, allowing it to keep context in long conversations. Nova 2 Omni Omni is a new addition to the Nova family. AWS describes it as its “first unified multimodal reasoning model.” It can handle text, images, video and audio as its inputs and create both text and image outputs. AWS says the model can handle the equivalent of 750,000 words of context (for some reason, AWS’s materials did not provide the number of tokens the model can handle). “While there are no comparable models in the industry to Nova 2 Omni, it demonstrates strengths in public benchmarks of multimodal reasoning on documents, images, videos, and audio, and can generate high-quality images similar to other leading image-generation models,” AWS says in its press materials about the new Omni model. The post AWS Updates Its Nova Models To Compete With Google, Anthropic and OpenAI appeared first on The New Stack.