Nova 2.5 Sonic is Amazon’s updated voice model, and it became generally available on Amazon Bedrock on October 5, 2026. If you build voice agents, it promises better reasoning and more accurate tool calls than Nova 2 Sonic, at the same price.
That last part matters most. You get a smarter model without a bigger bill, though Amazon hasn’t published numbers to prove the gains yet.
Here are the details worth knowing before you touch your code.
- Released: October 5, 2026, generally available on Amazon Bedrock
- Regions: N. Virginia, Oregon, Stockholm and Tokyo
- Context window: 256K tokens
- Price: same as Nova 2 Sonic
- Also new: Strands Bidi Agents framework is now generally available
Those five facts are the whole headline. The rest of this post explains what they mean in practice.
What Nova 2.5 Sonic actually does
It’s a speech-to-speech model. Audio goes in, spoken audio comes out, with no separate speech-to-text and text-to-speech steps stitched together. That’s how voice agents stay quick enough to feel like a conversation.
Amazon says this release improves reasoning, instruction following and tool-calling accuracy. It also says latency is lower. According to the AWS News Feed, the model is aimed at customer service, personal assistants and real-time conversational apps.
Voice and text can share one session. A user can speak a question and your app can still push typed context into the same conversation.
Async tool calling is the real upgrade
Picture a caller asking for an order lookup and a return. With older setups, the agent often went quiet while the lookup ran. Awkward silence makes people say “hello?” and hang up.
Asynchronous tool calling lets the agent keep talking while the tool works in the background. It can keep the caller company, then drop in the answer when it arrives. Amazon also says agents hold context across multi-step tasks like this.
The 256K-token context window helps too. Long support calls with plenty of history are less likely to push early details out of memory.
Where you can use it, and what we still don’t know
The model runs in four AWS regions for now. If your users or data rules require another region, you’ll have to wait. The table below separates what Amazon has stated from what the launch coverage left out.
| Topic | What’s stated | Not published yet |
|---|---|---|
| Reasoning and tool calls | Improved over Nova 2 Sonic | Benchmark scores |
| Languages | Seven, with expressive voices | The full list |
| Price | Same as Nova 2 Sonic | Dollar figures in the coverage I read |
| Latency | Lower than before | Actual milliseconds |
For the language list, the earlier model gives a clue. According to the AWS launch post for Nova 2 Sonic, that model covered English, French, Italian, German, Spanish, Portuguese and Hindi. That’s also seven, but check the Bedrock docs before you assume the lists match.
How to switch from Nova 2 Sonic
The original Nova 2 Sonic used the same bidirectional streaming API as the first Nova Sonic, so switching mostly meant changing the model ID. Amazon hasn’t said the same about 2.5 in the sources I checked, but it’s a sensible expectation.
Look up the exact 2.5 model ID in the Bedrock console rather than guessing it. Then test in a staging app before you move live traffic.
If you use the new Strands Bidi Agents framework, Amazon says Nova 2.5 Sonic works with minimal code. That’s the easiest on-ramp if you’re starting from scratch.
A short test plan helps you judge whether the upgrade is real for your use case. Run the same ten or twenty scripted calls through both models and compare tool-call accuracy and awkward pauses. Your own calls beat any launch claim.
Who benefits most? Teams running phone support, order-status lines or in-app voice assistants. If your agent mostly answers simple questions with no lookups, the async tool calling won’t change much for you, and the upgrade is lower priority.
How this fits the wider voice-AI push
Voice is getting attention everywhere right now. We covered how Codex CLI now turns voice on by default, which shows how quickly voice is becoming a normal way to talk to AI tools.
Model pricing is moving fast too. Our look at Claude Haiku 5.5’s API price cut is a good reminder to re-check your costs each quarter, not once a year.
Frequently Asked Questions
What is Nova 2.5 Sonic?
It’s Amazon’s speech-to-speech AI model for real-time voice agents, available through Amazon Bedrock. It handles spoken conversation and tool calls in one model.
Is Nova 2.5 Sonic free?
No. Amazon says pricing matches Nova 2 Sonic, so you pay Bedrock’s usage rates. Check the Bedrock pricing page for current figures.
Which regions support Nova 2.5 Sonic?
At launch it’s available in US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo).
Do I need to rewrite my app to upgrade?
Probably not much. The earlier Nova Sonic models shared one streaming API, so a model ID change is the likely main step. Test first to be sure.
What is Strands Bidi Agents?
It’s an AWS framework for building agents that talk both ways in real time. AWS made it generally available alongside this model.
My take: if you already run a voice agent on Nova 2 Sonic, test 2.5 this week. The price is the same, so the only cost is your testing time. If you’re choosing a voice model from scratch, wait for independent benchmarks before you commit, because “improved” with no numbers isn’t a buying reason yet.
Leave a Reply