Amazon has released Nova 2.5 Sonic, a new speech-to-speech model for real-time voice agents, through Amazon Bedrock. The model is generally available in four AWS regions and keeps the same pricing as Nova 2 Sonic, while Amazon claims better reasoning, instruction following, tool selection and response latency.
The more consequential part of the launch is not a more natural synthetic voice. Nova 2.5 Sonic is designed to stay inside a live conversation while it calls software tools, waits for results and completes multi-step work. Amazon’s example is a support agent that can find an order, check whether it qualifies for a return and start the return without handing the customer to a separate form.
That makes the release relevant to teams building call-center automation, travel assistants, healthcare scheduling, financial-service support or any other voice workflow that can change records. It also raises a harder deployment question: how do you keep a fast, conversational agent from taking the wrong action?
What Nova 2.5 Sonic adds
Nova 2.5 Sonic accepts and produces both speech and text. It supports seven languages, polyglot voices, controllable turn-taking and asynchronous tool calls, according to Amazon’s launch notice. An asynchronous call lets the agent continue managing the conversation while an order system, database or another agent works in the background.
The model is available in Amazon Bedrock in US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo). The model identifier used by the Strands framework is amazon.nova-2-5-sonic.
AWS has not published independent or directly comparable performance results for the 2.5 release. The announcement describes improved latency and tool-calling accuracy but does not provide error rates, latency percentiles or a benchmark table. Buyers should treat those improvements as vendor claims until they can test the model with their own accents, background noise, tools and business rules.
The context-window documentation does not agree
Teams planning long calls should verify the model limits in their own Bedrock account. Amazon’s October 5 announcement lists a 256,000-token context window. The current Nova product page and Nova 2 documentation, however, describe support for up to one million tokens.
Those figures may reflect different configurations, a staged update or documentation published out of sequence. Amazon does not explain the discrepancy on the cited pages. A fourfold difference can change how an application handles long conversations, retrieved documents and tool results, so developers should not size session memory from a marketing page alone.
There is a separate connection limit to consider. Strands says a Nova 2.5 Sonic connection lasts eight minutes. Its newly general-available Bidi Agents framework can restart the provider connection before that deadline and carry the conversation into a replacement connection. That makes a longer call possible, but it also creates a boundary that engineering teams need to test for dropped audio, repeated messages, tool calls in flight and lost conversational state.
Strands Bidi Agents turns the model into a working stack
Amazon paired the model release with general availability of Strands Bidi Agents, an open framework for bidirectional voice applications. It supports Nova Sonic, OpenAI Realtime and Gemini Live models behind a common agent interface.
The framework handles streaming microphone and speaker audio, user interruptions, acoustic echo cancellation, noise suppression and connection restarts. It can also expose tools or delegate a task to another agent while the voice model continues the front-end conversation.
That provider-neutral design matters for production systems. A team can keep its audio input, tool definitions and agent lifecycle while testing another model provider. Provider-specific voice and turn-detection settings still require work, but the application does not have to be rebuilt around a different orchestration library.
Strands also emits OpenTelemetry spans for sessions, model responses, tool calls, connections and restarts. Operators can measure time to first audio, tool latency and token use instead of debugging a failed call from the final transcript alone. Sensitive trace attributes can be redacted, but application logs and stored conversation history need their own retention and access controls.
AWS’s travel demo shows the real architecture
An AWS reference implementation published October 6 is more revealing than the short product announcement. It builds an airline concierge that can retrieve an itinerary, change a seat, update a meal preference, answer policy questions and escalate to a person.
The browser sends 16 kHz PCM audio over a signed WebSocket connection. Nova 2.5 Sonic runs inside Amazon Bedrock AgentCore, while AgentCore Gateway exposes backend functions as Model Context Protocol tools. API Gateway, Lambda and DynamoDB handle booking data and changes. A Bedrock Knowledge Base retrieves policy passages, and Cognito supplies user authentication and temporary AWS credentials.
This is a substantial cloud architecture, not a model dropped into a phone line. The sample uses per-session microVM isolation, signed connections, IAM-authorized APIs, encrypted storage, monitoring and separate front-end, agent and backend layers. Teams should include that operational footprint when comparing Nova 2.5 Sonic with a simpler transcription-plus-chat pipeline.
Five tests to run before letting a voice agent make changes
- Test tool selection and parameters separately. A pleasant response can hide a wrong account number, date or action. Score whether the agent chose the right tool and supplied every argument correctly.
- Require confirmation before writes. AWS’s travel sample asks the user to confirm before changing a booking. Apply the same pattern to payments, cancellations, address changes, account recovery and other consequential actions.
- Exercise the eight-minute handoff. Run calls across the connection restart while a tool is pending, while the user is speaking and immediately after confirmation. Confirm that no write runs twice.
- Measure tail latency. Average response time can look good while a slow database or retrieval call creates awkward silence. Track time to first audio, tool duration and end-to-end completion at the 95th and 99th percentiles.
- Audit transcripts, traces and audio retention. Voice sessions can contain payment data, health information, passwords and accidental background speech. Decide which layer stores each artifact, who can access it and when it is deleted.
Teams should also test accents, code-switching, noisy rooms, speakerphone echo, interruptions and ambiguous confirmations. The seven supported languages are useful only if the complete workflow, including speech recognition, tool arguments and spoken output, works reliably for the population being served.
Who should consider the upgrade
Nova 2.5 Sonic is most attractive to organizations already committed to Bedrock and building voice workflows that need tool use rather than simple question answering. Keeping the previous model’s price removes one adoption barrier, but Amazon’s public pricing page does not surface a clean 2.5-specific cost table. Teams should confirm regional Bedrock rates and model access before forecasting call costs.
Existing Nova 2 Sonic users have the clearest evaluation path: replay representative sessions against both versions and compare tool success, latency, interruptions and escalation rates. New projects should benchmark Nova against at least one competing real-time model and a conventional speech-to-text pipeline. Direct speech-to-speech can preserve tone and reduce conversational delay, while a modular pipeline can be easier to inspect, substitute and troubleshoot.
The release moves voice agents closer to completing useful transactions without breaking the conversation. Whether that becomes a better customer experience depends less on how human the voice sounds than on identity checks, tool permissions, confirmation design and the quality of the systems behind it.
