Amazon has added four new AI models to its Nova lineup, introduced a new way for companies to train their own custom versions, and launched a tool that helps build AI agents that can work inside web browsers, the company announced at re:Invent 2025 on Tuesday.
The company said tens of thousands of customers are already using Nova models for tasks including content generation, multi-step automation, and agent development. The new Nova 2 family is designed to balance speed, cost, and reasoning performance across text, image, video, and speech inputs.
New Nova 2 Models
Nova 2 Lite and Nova 2 Pro are built for reasoning-focused workloads with web grounding and code execution capabilities.
Nova 2 Lite is built for everyday applications such as customer support and document processing. Amazon said the model is “equal or better” across most benchmark comparisons with Claude Haiku 4.5, GPT-5 Mini, and Gemini Flash 2.5.
Nova 2 Pro, the company’s most capable reasoning model, is intended for complex tasks like agentic coding, long-range planning, and multi-document analysis. It outperformed or equalled Claude Sonnet 4.5, GPT-5.1, Gemini 2.5 Pro and Gemini 3 Pro Preview on a majority of benchmarks, according to Amazon.
Nova 2 Sonic is a speech-to-speech model built for real-time conversational AI with support for long context interactions and integration with telephony and voice frameworks.
Nova 2 Omni is a unified multimodal model that can process text, images, video, and audio while generating both text and images. Amazon said it can handle large-scale inputs such as product catalogues, long videos, and multi-format brand assets in a single workflow.
Organisations including Cisco, Siemens, Sumo Logic, and Trellix are using Nova 2 models for applications such as threat detection, video understanding, and voice assistants.
Nova Forge: Building Custom ‘Novella’ Models
Amazon also introduced Nova Forge, an open training capability that lets organisations build customised variants of Nova, called “Novellas,” by blending proprietary datasets with Nova’s training stages.
The service provides access to pre-trained, mid-trained, and post-trained checkpoints, allowing customers to integrate domain-specific knowledge throughout the training cycle. Amazon said the approach avoids the trade-offs of shallow fine-tuning or training from scratch.
Nova Forge includes reinforcement learning environments (“gyms”), support for synthetic data-driven distillation to create smaller models, and a responsible AI toolkit. Customers can deploy their custom models on Amazon Bedrock.
Early adopters include Booking.com, Reddit, Sony, Cosine AI, and Nomura Research Institute.
Nova Act: Automating Browser-Based Workflows
Amazon also launched Nova Act, a service for building and deploying AI agents that automate actions in web browsers. Powered by a Nova 2 Lite variant, Nova Act has reached 90% reliability in early customer workflows, the company said.
Nova Act uses reinforcement learning over thousands of simulated web tasks to improve its performance on UI-based actions such as CRM updates, website testing, and insurance form submissions.
Customers can prototype agents using natural language in a no-code playground, refine them in tools like VS Code, and deploy them through AWS. Hertz, Sola Systems, 1Password, and Amazon’s Project Kuiper team are among early users.
Amazon’s Leo satellite internet team also used Nova Act for test automation. The company said the system reduced test case creation from weeks of engineering effort to minutes.
The post Amazon Fires Back at OpenAI and Google with New Nova Models and Nova Forge appeared first on Analytics India Magazine.
