Amazon Nova 2.5 Sonic is a speech-to-speech model for developers building real-time, natural voice agents. Announced by AWS on October 5, 2026 as generally available, Nova 2.5 Sonic listens to spoken audio and replies in speech directly, without a separate transcribe-then-think-then-talk chain. It is the follow-up to the original Nova Sonic and is positioned for production contact centers rather than toy demos.
Core Features
- Unified speech understanding and generation with lower latency than the original Nova Sonic.
- Improved reasoning, instruction following, and tool-calling accuracy for multi-step customer tasks.
- Seven languages with expressive, polyglot voices and a 256K-token context window.
- Cross-modal sessions that switch between voice and text within the same conversation.
- Asynchronous tool use, so it can keep talking while a background task runs; works with Strands Bidi Agents on Bedrock.
Use Cases
- Customer-support voice agents that look up an order, check return eligibility, and start the return in one call.
- Interactive learning and voice-enabled personal assistants.
- Real-time conversational apps that must complete tasks, not just answer questions.
Pricing
Nova 2.5 Sonic is usage-based through Amazon Bedrock, billed per minute of audio at the same rate as the original Nova Sonic tier. It is available in US East, US West, Europe (Stockholm), and Asia Pacific (Tokyo), requires an AWS account, and has no free standalone tier.
Our Take
Best for production voice agents that must actually get things done rather than just chat; the trade-off is limited launch coverage (four regions) and no independent benchmarks yet to confirm the speed claims. See related models in our AI Audio tools category.
