Amazon Nova 2.5 Sonic: Features, Languages, Regions and How It Works

Quick answer: Amazon Nova 2.5 Sonic is AWS’s latest real-time speech-to-speech model for voice agents in Amazon Bedrock. AWS announced general availability on October 5, 2026. The update improves reasoning, instruction following, tool-calling accuracy and latency while keeping the conversational voice features introduced with Nova 2 Sonic.

Last checked: October 7, 2026.

What is Amazon Nova 2.5 Sonic?

Nova 2.5 Sonic is designed for applications where a user speaks naturally and expects the AI to understand, reason, use tools and respond with speech in real time. Unlike a pipeline that separately runs speech recognition, a text model and text-to-speech, Nova Sonic is built around direct conversational speech interaction.

AWS positions the model for customer-service agents, voice-enabled personal assistants, interactive learning products and other applications where low-latency spoken interaction matters.

What changed in Nova 2.5 Sonic?

AWS highlights four improvements over the previous generation:

• Stronger reasoning: the model is better at following multi-step requests during a conversation.
• Better instruction following: developers can steer agent behavior more reliably.
• More accurate tool calling: voice agents can use connected systems more effectively for tasks such as looking up an order, checking eligibility or starting a workflow.
• Lower latency: faster responses make the conversation feel less like a sequence of disconnected turns.

Supported languages

Amazon says the current Sonic family supports seven languages with expressive voices and automatic language switching:

• English
• French
• Italian
• German
• Spanish
• Portuguese
• Hindi

English voice support spans multiple accents, including U.S., UK, India and Australia. Polyglot voices are designed to keep a consistent voice identity even when the speaker changes languages within the same session.

Voice and text can share the same session

Nova 2.5 Sonic supports cross-modal interaction, so an application can move between voice and text without treating them as completely separate conversations. That is useful for agents that need to speak with a customer but also display a confirmation, reference number or written summary.

How tool calling works in a voice agent

The model can call external tools while the conversation continues. A customer-support example might work like this:

1. The customer asks about an order by voice.
2. The agent uses a connected order system to find it.
3. The user asks whether it can be returned.
4. The agent checks a return-policy or eligibility tool.
5. If eligible, the agent starts the return workflow.
6. The voice conversation continues while those operations happen.

AWS describes support for asynchronous tool calling, which is especially useful when a tool takes longer than a normal conversational turn.

Context window

AWS’s October 5 announcement lists a 256K context window for Nova 2.5 Sonic. That capacity is designed to support longer-running voice sessions and substantial conversational context without implying the 1-million-token context available on some other Nova models.

Turn-taking and natural conversation

Real-time voice agents have to decide when the user has finished speaking and when the model should respond. The Sonic family includes controllable turn-taking and pause sensitivity so developers can tune how quickly the system reacts.

This is important because an agent that interrupts too aggressively feels unnatural, while one that waits too long feels slow even if the model itself is fast.

Where is Nova 2.5 Sonic available?

AWS says Nova 2.5 Sonic is available through Amazon Bedrock in these Regions:

• US East (N. Virginia)
• US West (Oregon)
• Europe (Stockholm)
• Asia Pacific (Tokyo)

Region availability can change, so production deployments should verify the current Amazon Bedrock model-region table before choosing an architecture.

How much does Nova 2.5 Sonic cost?

AWS says Nova 2.5 Sonic is offered at the same pricing as Nova 2 Sonic. Because Bedrock pricing can change and may differ by usage type or region, developers should use the live Amazon Bedrock pricing page for the exact current rate rather than relying on a static number copied into an article.

How is Nova 2.5 Sonic different from Nova 2 Sonic?

Strands Bidi Agents integration

AWS also points developers to Strands Bidi Agents for building production voice agents with Nova 2.5 Sonic. The integration is intended to reduce the amount of low-level plumbing needed for bidirectional voice conversations and tool-connected workflows.

Best use cases

• Customer service: order lookups, policy checks, account actions and guided support.
• Personal assistants: voice-first task completion across connected services.
• Interactive learning: spoken tutoring and practice sessions.
• Contact-center automation: conversational agents that can use enterprise tools.
• Voice interfaces for apps: natural speech control without building a separate ASR + LLM + TTS stack.

What developers should verify before production

Before shipping a voice agent, test language quality, interruption behavior, tool failures, long pauses, noisy environments, authentication, data retention and fallback behavior. A voice system can feel highly capable in a demo but still fail if a backend tool times out or the user changes topic mid-task.

FAQ

Is Nova 2.5 Sonic generally available?

Yes. AWS announced general availability on October 5, 2026.

Is it text-to-speech only?

No. It is a speech-to-speech model designed for real-time conversations, with voice and text able to coexist in a session.

Does it support tool use?

Yes. AWS specifically highlights improved tool-calling accuracy and asynchronous tool calling.

How many languages does it support?

AWS lists seven languages: English, French, Italian, German, Spanish, Portuguese and Hindi.

Official sources

• AWS announcement: Nova 2.5 Sonic
• Amazon Nova models
• Amazon Nova speech-to-speech documentation