Single-Model Design Eliminates Pipeline Latency

Instead of chaining separate speech-to-text, reasoning, and text-to-speech services, Nova 2 Sonic unifies speech understanding and generation into one model. This design is what enables low-latency two-way voice, making conversations feel more natural and responsive.

Business outcome: Reduces conversational lag, leading to higher customer satisfaction in real-time voice interactions.

Per-Token Pricing Cuts Cost Per Minute Dramatically

Independent reporting indicates Nova 2 Sonic costs roughly $0.015 per minute, which is around 80% cheaper than the OpenAI GPT-4o Realtime alternative. Speech is priced at $0.003 per 1,000 input tokens and $0.012 per 1,000 output tokens.

Business outcome: Significantly lowers the operational cost of running AI voice agents at scale, making 24/7 availability more viable.

Seven Languages with Polyglot Voice Support

The model supports seven languages and includes polyglot voices, allowing a single voice to switch between languages mid-conversation. This is a practical feature for global customer service operations that serve multilingual audiences.

Business outcome: Enables a single voice agent to serve customers in multiple languages without complex routing or separate voice models.

Cross-Modal Interaction for Flexible Sessions

A session can switch between voice and text interaction, meaning a conversation that starts as a voice call can transition to text without losing context. This flexibility supports omnichannel customer journeys within a single session.

Business outcome: Provides seamless customer experiences across channels, reducing friction when users move from voice to chat.

1M-Token Context Window for Long Conversations

With a context window of up to 1 million tokens, the model can retain and reference a large amount of conversation history. This is particularly useful for complex support scenarios where context from earlier in the call matters.

Business outcome: Maintains coherence in long-running interactions, reducing the need for customers to repeat information.

Native Availability on Amazon Bedrock

Nova 2 Sonic is available through Amazon Bedrock in US East (N. Virginia), US West (Oregon), and Asia Pacific (Tokyo). For teams already using AWS, this means straightforward integration with existing security, monitoring, and deployment tooling.

Business outcome: Accelerates time-to-deployment for AWS-centric teams by removing the need to manage separate infrastructure.

• Cost EfficiencyIndependent reporting places the cost at around $0.015 per minute, roughly 80% cheaper than OpenAI GPT-4o Realtime, making it a clear economic choice for high-volume voice traffic.
• Low Latency ArchitectureThe single-model speech-to-speech design avoids the latency of chained systems, which is essential for natural real-time conversation.
• Cross-Modal FlexibilityThe ability to switch between voice and text within a session offers operational flexibility that many dedicated voice or text models lack.
• Long Context HandlingA 1M-token context window supports complex, long-running interactions without losing track of earlier conversation details.

• Regional AvailabilityLimited to US East (N. Virginia), US West (Oregon), and Asia Pacific (Tokyo), which rules out deployments requiring data residency in the EU or other regions.
• AWS Ecosystem DependencyOnly available through Amazon Bedrock, meaning teams not already on AWS face a platform commitment to use this model.
• No Free TierThere is no free tier or trial; usage requires a paid AWS account, which can complicate initial evaluation for smaller teams.
• Professional RealityWhile the per-minute cost is compelling, the hard regional lock-in is a dealbreaker for any organisation with strict data sovereignty requirements outside the three supported AWS regions.

High-Volume Contact Centres

For operations handling thousands of calls daily, the low per-minute cost of Nova 2 Sonic makes AI-powered handling of routine queries economically viable, freeing human agents for complex issues.

Real-Time Voice Assistants

Product teams building voice assistants where response time is critical will benefit from the unified model architecture that minimises conversational lag.

Multilingual Customer Support

With seven languages and polyglot voices, businesses serving diverse language bases can deploy a single system that handles multiple languages without separate models.

AWS-Native Voice Applications

Enterprises already running on Amazon Bedrock can integrate voice AI with their existing cloud stack, using the same security and monitoring frameworks they already trust.

Set up an AWS account and navigate to the Amazon Bedrock console to request access to the Nova 2 Sonic model.

Review the model's pricing page and estimate your expected token usage based on call volume and average conversation length.

Use the AWS Bedrock API documentation to integrate the speech-to-speech model into your existing application architecture.

Run a pilot with a small set of test conversations to measure latency and conversation quality before scaling to production traffic.