Amazon Nova Model Family Upgrade: Stronger AI, Lower Latency, Better Cost‑Performance

At re:Invent 2025 Amazon Web Services announced a major expansion of the Amazon Nova product line, adding four new models, a groundbreaking “open‑training” service for custom model creation, and a high‑reliability AI Agent offering.

Amazon Nova 2 Lite is a fast, economical inference model for everyday workloads that handles text, image, and video inputs and lets users adjust the model’s “thinking” depth to balance intelligence, latency, and cost. In head‑to‑head benchmarks it matches or exceeds Claude Haiku 4.5 on 13 of 15 tests, GPT‑5 Mini on 11 of 17, and Gemini Flash 2.5 on 14 of 18.

Amazon Nova 2 Pro is the most intelligent inference model, supporting text, image, video, and voice inputs for highly complex tasks such as agent programming, long‑term planning, and advanced problem solving. It can act as a teacher model for knowledge distillation. Compared with Claude Sonnet 4.5 it is equal or better on 10 of 16 evaluations, with GPT‑5.1 on 8 of 16, Gemini 2.5 Pro on 15 of 19, and Gemini 3 Pro Preview on 8 of 18. It excels in multi‑document analysis, video reasoning, complex instruction execution, high‑level mathematics, and software‑engineering tasks.

Amazon Nova 2 Sonic delivers end‑to‑end speech capabilities, merging text and voice understanding and generation for real‑time, human‑like conversations. It supports many languages, expressive voice tones, and up to 1 million‑token context windows, enabling long‑duration interactions and asynchronous task handling. It integrates seamlessly with Amazon Connect, third‑party voice providers (Vonage, Twilio, AudioCodes) and conversational AI frameworks (LiveKit, Pipecat). Compared with OpenAI gpt‑realtime and Gemini 2.5 Flash, Sonic leads in cost‑performance and voice quality.

Amazon Nova 2 Omni is a unified multimodal model that processes text, image, video, and audio while simultaneously generating text and images—a first in the industry. It can handle up to 750 k words of text, hours of audio, long video, and hundreds of pages of documents in a single pass, allowing a workflow that instantly creates complete marketing assets (titles, copy, social posts, visual designs). Public multimodal benchmarks show it outperforms peers and produces images comparable to leading generators.

Amazon Nova Forge addresses three unsatisfactory options for infusing proprietary knowledge into AI: limited fine‑tuning of closed models, open‑weight training without original data (risking capability degradation), and building from scratch (high cost and time). Forge offers an “open‑training” path by releasing pre‑training, mid‑training, and post‑training checkpoints, letting customers mix their data with curated Amazon Nova datasets to create custom “Novellas.” It also provides three key capabilities: (1) a custom reinforcement‑learning “gym” for domain‑specific simulation environments, (2) a distillation pipeline to produce smaller, faster high‑performance models, and (3) a responsible‑AI toolkit for security, compliance, and governance.

Amazon Nova Act is a high‑reliability AI Agent service for browser‑based UI workflows, powered by a customized Nova 2 Lite. It achieves 90 % execution reliability and outperforms competing agents in benchmark tests. Trained via reinforcement learning on hundreds of simulated web environments, it excels at tasks such as CRM data updates, website testing, and insurance claim submissions. Developers can prototype agents in a zero‑code visual environment, refine them in familiar IDEs (e.g., VS Code), and deploy at scale on Amazon Bedrock with enterprise‑grade security, scalability, and data‑privacy guarantees.

Enterprise customers—including Sola Systems, 1Password, Hertz, Amazon Leo, Reddit, Booking.com, Cosine AI, Nimbus Therapeutics, and Sony—have already built custom models or agents with Nova Forge/Act, reporting massive automation gains, faster testing cycles, and reduced AI costs. Deployments on Amazon Bedrock ensure consistent security, scalability, and privacy across production workloads.

For more details, see the Amazon Nova product page and the developer guide linked in the original article.