Amazon Nova 2.5 Sonic is AWS’s speech-to-speech model for real-time voice agents, announced as generally available on October 5, 2026 in Amazon Bedrock. It is a point release of Nova 2 Sonic (December 2025), and AWS is explicit about what changed: better reasoning, instruction following and tool-calling accuracy, and lower latency — with the same API surface, voices, language support and price.
That framing matters. This is not a new model family or a new price tier; it is a smarter brain for the same voice stack. If you already build on Nova 2 Sonic, the upgrade path is a model change rather than a rewrite.
What changed from Nova 2 Sonic
- Reasoning and instruction following. The model is meant to hold context, follow multi-step instructions and use connected tools to finish customer tasks — AWS’s example is an agent that looks up an order, checks return eligibility and starts the return during the call.
- Tool-calling accuracy. The service card says open-source models were used to improve tool calling and instruction following.
- Lower latency for faster, more natural turn-taking. AWS gives no figure.
- Unchanged: the same API surface, the same 18 voices and the same language support as Nova 2 Sonic.
Capabilities
- Speech understanding that is robust to varied speaking styles, with adaptive responses that follow the prosody of the user’s speech.
- Expressive speech generation in 18 voices, masculine- and feminine-sounding, across English (US, British, Indian and Australian), Spanish, German, French, Italian, Brazilian Portuguese and Hindi.
- Controllable turn-taking and graceful handling of interruptions without dropping context.
- Voice and text in the same session, with asynchronous tool calling so the agent can keep talking while a task runs.
- 256K context window per AWS’s announcement.
- Cross-modal input and agent frameworks: Nova 2.5 Sonic works with Strands Bidi Agents, now generally available, which handle audio I/O, tools, longer conversations and connections to other agents in a few lines of code, with OpenTelemetry tracing.
Pricing
AWS says Nova 2.5 Sonic is priced the same as Nova 2 Sonic. AWS’s own pricing page renders the numbers dynamically, so the per-token rates below are the standard-tier figures third-party indexes track for Nova 2 Sonic. Confirm them in the Bedrock console:
| Token type | Per 1M tokens |
|---|---|
| Speech input | $3.00 |
| Speech output | $12.00 |
| Text input | $0.33 |
| Text output | $2.75 |
Text-token rates apply to transcription, tool calls, grounding and conversation history. Regional pricing varies — Tokyo lists higher rates than US East. Bedrock also offers Priority, Flex and Reserved service tiers on some models; check which tiers apply to Nova 2.5 Sonic.
Availability and limits
- Regions: Amazon Bedrock in US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo).
- Model ID: Strands Agents’ docs use
amazon.nova-2-5-sonic. - Connection length: each Nova Sonic connection is limited to eight minutes. The Strands
BidiAgentschedules a restart after seven minutes and replays up to 200 KiB of recent history into the new connection. - Audio: mono PCM; Strands defaults to 16 kHz in and out and lets you set 24 kHz output.
- Language scope: officially released and supported for English, Spanish, German, French, Italian, Portuguese and Hindi only. Other languages may give reduced speech recognition accuracy, less natural voices or content errors.
- Context discrepancy: AWS’s announcement lists a 256K context window; the Bedrock model card for Nova 2 Sonic lists 1M tokens. Treat the exact limit as something to verify in your region.
What AWS has not published
There are no benchmark scores for Nova 2.5 Sonic, no latency numbers and no head-to-head against GPT-class realtime models or Gemini Live. The service card describes the evaluation approach and intended uses, but not results. Until independent measurements appear, treat “improved reasoning” as a vendor direction, not a measured gain.
How it compares
- vs Qwen-Audio 3.1 Realtime: Alibaba’s full-duplex model covers 11 languages plus Chinese varieties, publishes benchmark results (τ²-Bench Audio 82.19%) and lists audio output at $24 per 1M tokens; Nova 2.5 Sonic wins on AWS integration, Bedrock security and billing, and published pricing continuity.
- vs other realtime voice models: Strands’ launch post lists GPT-Realtime-2.1 and Gemini 3.8 Live as the other frontier voice models Bidi Agents connect to — pick on language coverage, tool-calling accuracy in your own dialogues and cloud lock-in.
- vs MAI-Voice-2.1: MAI-Voice is a text-to-speech model you pair with a separate transcription model; Nova Sonic is a single speech-to-speech model that listens, reasons and speaks in one stream.
- vs Gemini 3.8 Flash TTS: Gemini Flash TTS is a non-conversational speech generator; Nova Sonic is built for live, interruptible, tool-using dialogue.
Limitations
- No benchmarks or latency figures for 2.5.
- Seven supported languages and four regions.
- Eight-minute connection limit requiring restarts and history replay.
- Pricing details need confirmation on AWS’s dynamic pricing page.
See the audio models directory for other speech and voice models.
Best for
- Customer-service voice agents that look up records and take actions mid-call
- Personal voice assistants and real-time conversational apps
- Voice interfaces on AWS-native stacks that need Bedrock security and billing
- Agent-to-agent and multi-step workflows using Strands Bidi Agents
Pros & cons
Strengths
- Targets the hard part of voice agents — multi-step instructions, context retention and tool calls that complete customer tasks (look up an order, check return eligibility, start a return) inside one conversation
- Lower latency than Nova 2 Sonic, with controllable turn-taking, voice and text in the same session and asynchronous tool calling
- Drop-in successor — same API surface, voices and language support as Nova 2 Sonic, and the same pricing
- Production path through Amazon Bedrock plus Strands Bidi Agents (now generally available), which also work with other providers' voice models
Limitations
- No published benchmark numbers for 2.5 — AWS cites improvements in reasoning, instruction following and tool-calling accuracy without scores
- Officially supports seven languages only (English, Spanish, German, French, Italian, Portuguese, Hindi); AWS advises against other languages for now
- Regions are limited to four (N. Virginia, Oregon, Stockholm, Tokyo); each Bedrock Sonic connection is limited to eight minutes, so long calls need restarts
- AWS lists a 256K context window for 2.5, while Nova 2 Sonic's model card lists 1M — confirm limits in the Bedrock console before designing around either number
Frequently asked questions
What is Amazon Nova 2.5 Sonic?
Amazon Nova 2.5 Sonic is AWS's speech-to-speech model for natural, real-time voice agents, generally available since October 5, 2026 on Amazon Bedrock. It succeeds Nova 2 Sonic with improved reasoning, instruction following and tool-calling accuracy, and lower latency.
How much does Nova 2.5 Sonic cost?
AWS says it costs the same as Nova 2 Sonic. For Nova 2 Sonic, third-party trackers list standard-tier rates of $3 per 1M speech input tokens and $12 per 1M speech output tokens, plus $0.33 and $2.75 per 1M text input and output tokens for transcription, tool calls and history. Verify current rates and regional differences on the Amazon Bedrock pricing page.
What languages does Nova 2.5 Sonic support?
English (US, British, Indian and Australian), Spanish, German, French, Italian, Brazilian Portuguese and Hindi, with 18 expressive voices. AWS says the model has been trained on multilingual data but does not recommend other languages, which may reduce recognition accuracy and naturalness.
Is Nova 2.5 Sonic better than Nova 2 Sonic?
AWS says it brings stronger reasoning, instruction following and tool-calling accuracy and lower latency, and its service card says open-source models were used to improve tool calling and instruction following. AWS has not published benchmark numbers, so test it on your own tool-using dialogues before migrating.
Where is Nova 2.5 Sonic available?
Amazon Bedrock in US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo). Strands Agents documents the model ID as amazon.nova-2-5-sonic for use with its Bidi Agents.
What is the connection limit?
Strands Agents documents that Nova Sonic limits each connection to eight minutes. Its BidiAgent restarts the connection after seven minutes by default and replays conversation history into the new one.