Amazon unveils four Nova AI models for automating tasks

Amazon has launched four new Nova AI models, a custom model-building service called Nova Forge, and a new agent-building tool named Nova Act.

The Nova 2 lineup includes Lite and Pro models for reasoning, Sonic for speech-to-speech, and Omni for multimodal tasks across text, images, video, and speech.

Nova Forge lets organizations create tailored versions of Nova models using their own data at multiple training stages, aiming to address demands for deeper business integration.

Nova Act enables users to build AI agents that automate browser-based tasks, with Amazon reporting early customer workflows achieving 90% reliability.

All new Nova services are available on AWS, allowing customers to deploy custom models with enterprise security and privacy features.

🔗 Source: Amazon

🧠 Food for thought

Implications, context, and why it matters.

Nova 2 Lite tops rivals on tests, yet AWS omits latency and cost data

Tests put Nova 2 Lite at or above Claude Haiku 4.5 (Anthropic), GPT-5 Mini (OpenAI), and Gemini Flash 2.5 (Google) on several benchmarks 1.
AWS leaves out head-to-head latency comparisons and total cost-of-ownership metrics 2. Those should factor context window use (how much input text a model can handle at once), thinking budgets (settings that allocate more compute for reasoning), and tool execution costs.
Nova 2 Lite offers three thinking budget levels (low, medium, high) 2. AWS provides no benchmarks for these modes. AWS says Amazon Nova Premier generates up to 190 fewer tokens (fragments of words) per answer than comparable models 3. Similar data for Nova 2 Lite modes or multimodal inputs is missing.

Integrators can package Nova Act agents with observability and human-in-the-loop (HITL) patterns

For systems integrators, consultancies that implement and run complex software for large enterprises, Nova Act supports human-in-the-loop escalation for unresolved tasks 4. That can drive services work to shore up reliability in real workflows.
The platform includes act traces (step-by-step execution logs) and session recording 4. Teams can store session artifacts in Amazon S3 (AWS’s cloud object storage service). They can build observability (monitoring and debugging visibility) dashboards and debugging tools.
Integrators can offer Nova Forge custom model development 5. This helps enterprises bring in proprietary data during pre-training (broad learning from large datasets), supervised fine-tuning (training on labeled examples), and alignment (instruction-following and safety tuning).