AWS launches Nova 2.5 Sonic for real-time voice agents

AWS announced Nova 2.5 Sonic as a speech-to-speech model for real-time voice agents, and made it generally available on October 5, 2026.[1] AWS says the model improves reasoning and instruction following in real-time voice applications, as well as tool-calling accuracy.[1] It also says lower latency enables faster interactions.[1] The intended applications AWS names include customer service, personal AI assistants and real-time conversational applications.[1]

AWS says Nova 2.5 Sonic is intended to help voice agents maintain context and follow multi-step instructions.[1] Agents can use connected tools with the model to complete customer tasks, according to AWS.[1] The model retains speech-understanding capabilities and offers expressive voices in seven languages.[1] It also includes controllable turn-taking, supports voice and text in the same session, and allows asynchronous tool calling.[1] Those capabilities give teams several design choices to assess when deciding how an agent should handle spoken turns, text and tool use.

AWS says Nova 2.5 Sonic can be used with Strands Bidi Agents, which are now generally available.[1] AWS says developers can use Strands Bidi Agents to build production voice agents in a few lines of code.[1] The agents also support longer conversations and connections to other agents, according to AWS.[1] Nova 2.5 Sonic has a 256K context window.[1]

Nova 2.5 Sonic is available through Amazon Bedrock in four AWS Regions: US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo).[1] AWS says the model costs the same as Nova 2 Sonic.[1] The listed regions provide a concrete starting point for checking whether the service is available in a team's intended deployment location.