Aws Updates Its Nova Models To Compete With Google, Anthropic And Openai

LAS VEGAS — At its re:Invent convention successful Las Vegas today, AWS launched nan latest procreation of its family of Nova instauration models. Nova 2 Lite, Pro and Sonic — arsenic good arsenic a caller Omni exemplary that’s caller to nan family and that tin return text, audio, images and video arsenic inputs and output matter and images.

AWS announced nan first group of Nova models astatine past year’s re:Invent and released a number of further models for image generation, video procreation and different usage cases complete time. For nan astir part, though, these models ne'er rather achieved nan mindshare of models from different vendors and by now, those original models person fallen good down nan state-of-the-art models from frontier labs for illustration Anthropic, OpenAI and Google — arsenic good arsenic galore open-weight models.

With this caller release, AWS is catching up, though, and nan institution isn’t acrophobic to mention benchmarks that put these models up against existent frontier models (many of which only launched successful nan past fewer weeks).

Nova 1 Lite

The Lite version, for example, which AWS argues is ideally suited for usage successful customer work scenarios, archive processing and workflow automation, is competitory and sometimes outperforms models for illustration Anthropic’s Claude Haiku 4.5 and OpenAI’s GPT-5-mini (and Google’s Gemini Flash 2.5, but that’s an older exemplary astatine this point).

The halfway accuracy down this model, akin to Anthropic’s Haiku family, is to find nan correct tradeoff betwixt velocity and cost.

The Nova 2 Lite exemplary besides powers AWS’s caller Nova Act service, its caller work for managing browser-use agents. AWS says Nova Act delivers 90% reliability connected nan customer workflows it tested.

Nova 2 Pro

Nova Pro is AWS’s astir intelligent reasoning model, nan institution argues, that excels astatine tasks that request accuracy, multidocument analysis, reasoning complete images and video, and coding tasks. The benchmarks put it successful a convention pinch Claude Sonnet 4.5 and GPT-5- and GPT-5.1, arsenic good arsenic Google’s Gemini 3 preview (with Gemini 3 only launching past week, AWS did not supply benchmarks against nan last type of Gemini 3 yet).

It would beryllium bully to spot a comparison pinch Anthropic’s Opus 4.5, since that hit Gemini 3 Pro connected galore benchmarks, but it’s worthy noting that Nova 2 Pro only hit aliases matched Gemini 3 Pro connected 8 retired of 18 benchmarks.

As AWS notes, November 2 Pro tin besides beryllium utilized arsenic a coach exemplary for knowledge distillation.

Nova 2 Sonic

Sonic is simply a spot of a specialty model. It’s a speech-to-speech exemplary that’s astir apt champion suited for interaction centre solutions for illustration Amazon’s ain Connect work and different AI adjunct usage cases.

It’s meant for generating human-like conversations successful existent clip and successful aggregate languages pinch expressive voices. It tin move betwixt sound and matter arsenic needed and its token model is 1 cardinal tokens, allowing it to support discourse successful agelong conversations.

Nova 2 Omni

Omni is simply a caller summation to nan Nova family. AWS describes it arsenic its “first unified multimodal reasoning model.” It tin grip text, images, video and audio arsenic its inputs and create some matter and image outputs.

AWS says nan exemplary tin grip nan balanced of 750,000 words of discourse (for immoderate reason, AWS’s materials did not supply nan number of tokens nan exemplary tin handle).

“While location are nary comparable models successful nan manufacture to Nova 2 Omni, it demonstrates strengths successful nationalist benchmarks of multimodal reasoning connected documents, images, videos, and audio, and tin make high-quality images akin to different starring image-generation models,” AWS says successful its property materials astir nan caller Omni model.

Group Created pinch Sketch.