Amazon released Nova 2.5 Sonic on 5 October 2026 in four AWS Regions: US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo). It is the upgrade for teams already on Nova 2 Sonic who want better tool-calling accuracy and lower latency in call-center style voice agents.
What changed
- · Release Amazon Nova 2.5 Sonic generally available on Amazon Bedrock source
Amazon Nova 2.5 Sonic is Amazon's speech-to-speech model on Bedrock, generally available since 5 October 2026. It takes voice and text in the same session, speaks in seven languages with a 256K context window, and adds asynchronous tool calling for voice agents.
Provider: Amazon · Family: Nova 2.5
Context window: 256,000 tokens
Input modalities: text, audio, tool-calls · Output: text, audio, tool-calls
About Amazon Nova 2.5 Sonic
Amazon Nova 2.5 Sonic is a speech-to-speech model from Amazon, announced as generally available on Amazon Bedrock on 5 October 2026. Amazon describes it as the latest model in the Nova Sonic line for natural, real-time voice agents, following Nova 2 Sonic, which reached general availability in December 2025. It is a proprietary, API-only model: there are no downloadable weights, and Amazon has not published a parameter count or an architecture description.
The announcement lists the changes in plain terms. Amazon says this release brings improved reasoning, instruction following and tool-calling accuracy to real-time voice applications, together with lower latency for faster, more natural turns. It builds on existing Nova Sonic capabilities: speech understanding, expressive voices in seven languages, controllable turn-taking, voice and text in the same session, asynchronous tool calling and a 256K context window. Asynchronous tool calling means an agent can start a lookup while the conversation continues, and Amazon's own example is a customer-service agent that finds an order, checks return eligibility and starts a return within one conversation. Amazon has not published benchmark scores for the new model, so the reasoning and latency gains are the vendor's description and are not yet independently measured.
Nova 2.5 Sonic runs in Amazon Bedrock in four AWS Regions: US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo). Amazon states it is offered at the same pricing as Nova 2 Sonic. The release post does not repeat the rates and points to the Bedrock pricing page, whose model tables could not be read in full when this page was written, so HokAI lists no per-token price until the figures can be confirmed. Developers can also build with it through Strands Bidi Agents, which Amazon says is now generally available for production voice agents with longer conversations and connections to other agents.
It suits teams that already build on AWS and want a single model for the whole voice turn instead of chaining speech recognition, a language model and speech synthesis. Customer service, personal assistants and other real-time conversational products are the uses Amazon names. Teams that need a Region outside the four above, or a model they can host themselves, should look elsewhere, and so should anyone who needs published latency numbers before committing. Voice platforms such as Vapi and Retell AI assemble voice agents from interchangeable models, ElevenLabs is a speech-first alternative for synthesis, and Deepgram covers speech recognition. For text and multimodal reasoning on the same cloud, see Nova 2 Pro, and for another large vendor's offering compare OpenAI and Google DeepMind.
Pricing
Amazon says Nova 2.5 Sonic is priced the same as Nova 2 Sonic on Amazon Bedrock and refers readers to the Bedrock pricing page. Speech is metered separately from text tokens, and the rates were not readable from that page when this record was written, so no figure is stored. Check the live Bedrock pricing page for your Region before budgeting.
Key Features
- Speech-to-speech in one model: Handles spoken input and spoken output directly, with controllable turn-taking for natural interruptions.
- Voice and text in one session: A single session can mix spoken audio and typed text.
- Asynchronous tool calling: The agent can start a tool call, such as an order lookup, and keep talking until the result returns.
- Long single-session context: Holds long conversations and attached history without splitting the call.
Pros
- One model covers the spoken turn, which removes the speech-to-text and text-to-speech hand-offs.
- Asynchronous tool calling avoids dead air during lookups.
- Same pricing as Nova 2 Sonic, so an upgrade does not change the rate card.
Cons
- No published benchmarks or latency numbers.
- Four Regions only at launch.
- Bedrock-only, with closed weights.
Frequently Asked Questions
What does Nova 2.5 Sonic cost?
Amazon says it is offered at the same pricing as Nova 2 Sonic on Amazon Bedrock and does not restate the rates in the launch post. Speech and text are metered separately, so read the Bedrock pricing page for your Region before you size a budget.
Is Nova 2.5 Sonic better than Nova 2 Sonic?
Amazon says it improves reasoning, instruction following and tool-calling accuracy and lowers latency. No scores were published, so the gain is a vendor claim. Run your own recorded calls through both models if the decision matters.
Can you run Nova 2.5 Sonic yourself?
No. It is a proprietary model offered only through Amazon Bedrock, with no downloadable weights. Availability is limited to US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo).
Does Nova 2.5 Sonic train on your data?
The launch post does not say. AWS publishes its Bedrock data handling terms separately, so confirm retention and training terms there before sending customer calls to the model.
Who should pick Nova 2.5 Sonic?
AWS-native teams building voice agents that look up records mid-call fit best. A team that must host the model itself, or serve callers far from the four Regions, should compare Vapi or Retell AI, which let you swap speech models.