Amazon Nova 2.5 Sonic: a voice agent that can take the return call
Amazon Nova 2.5 Sonic is GA with better reasoning and tool calling for real-time voice agents, at Nova 2 Sonic pricing. Start with one call type, not your whole phone line.
AWS made Amazon Nova 2.5 Sonic generally available on October 5. It is a speech-to-speech model for real-time voice agents, and the upgrade is in the parts that break phone bots: reasoning, following multi-step instructions, and calling tools correctly mid-conversation. AWS's own example is the call every store gets: look up the order, then process the return, in one conversation.
What actually happened
Per AWS's announcement:
- GA now. Available in US East (N. Virginia), US West (Oregon), Europe (Stockholm), and Asia Pacific (Tokyo).
- Better at work, not just talk. Improved reasoning, instruction following, and tool-calling accuracy, plus lower latency.
- Same price as Nova 2 Sonic. No increase for the upgrade. Check the Bedrock pricing page for current per-token rates, since speech and text tokens bill differently.
- Seven languages and asynchronous tool calling carry over from Nova 2 Sonic, so the agent can keep talking while a lookup runs.
- Strands Bidi Agents is GA. AWS's open-source framework for two-way streaming agents, pitched as the short path to a production voice agent.
Why it matters for your business
Speech-to-speech means one model hears, thinks, and answers. The old stack chained speech-to-text, an LLM, and text-to-speech, and every hop added delay and a place to lose context. Fewer hops is why these agents finally sound like a conversation and not a phone tree.
The risk did not go away. It moved to the tools. An agent that can "process returns" is calling your order system with real permissions. So the first build should be narrow:
- One call type. Order status, or returns inside your policy window. Not "answer the phone."
- Read-only first. Let it look up orders for two weeks. Read the transcripts. Then give it the one write action it needs.
- Hard limits in the API. The return endpoint enforces the window and the dollar cap. Not the prompt.
- A clean handoff. Any doubt, it transfers to a human with the order number already pulled.
Count your calls before you build. If 40% of your inbound volume is "where is my order," that is the agent. The rest can wait.
Key takeaways
- Nova 2.5 Sonic is GA in four AWS regions at the same price as Nova 2 Sonic
- The upgrade targets reasoning, instruction following, and tool-calling accuracy
- Speech-to-speech cuts the delay of chained speech, LLM, and voice models
- Start with one call type and read-only tools, then add a single write action
- Enforce return windows and dollar limits in your API, not in the prompt
The phone call that eats your afternoon is usually the same one, forty times. We build voice agents that handle that one call, with scoped tools and a real handoff to your team. Price the hours it would save or tell us what your phone line sounds like.
Sources: AWS What's New, Amazon Nova pricing.
- #amazon-nova
- #voice-agents
- #aws-bedrock
- #customer-service
- #speech-to-speech
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Wajo AI agent calls businesses for customers: answer the bot
Khosla-backed Wajo's AI agent calls businesses and pays with virtual cards for its users. Here's how to get your phone line and booking flow ready.
Read itSAP Pay and Joule Work: your ERP now pays the invoice itself
SAP Pay executes and reconciles invoice payments inside SAP Cloud ERP, and Joule Work starts its customer rollout. Tighten approvals before AI touches cash.
Read it