Tasks and inputs
Nova 2.5 Sonic is listed for Audio-to-audio, Text → Speech, Speech → Text. It accepts speech, text, audio and produces speech, text. Use the publisher’s input format and preprocessing for your chosen deployment.
Nova 2.5 Sonic is Amazon's speech-to-speech model on Bedrock for real-time voice agents, with async tool calling, polyglot voices and a 256K context.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Pricing published on Amazon Bedrock's pricing page; not verified here. Standard tier only. Confirm on your own account.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Amazon's voice-agent model lives behind one Bedrock operation, InvokeModelWithBidirectionalStream. You will not reach it through Converse, Invoke or the Chat Completions-style endpoints. That single constraint shapes your whole architecture, so it belongs at the top of any evaluation.
Nova 2.5 Sonic launched on 5 October 2026 as the successor to Nova 2 Sonic. Amazon describes better reasoning, instruction following and tool-calling accuracy, together with lower latency and shorter, more concise replies.
| Model ID | amazon.nova-2-5-sonic |
| Endpoint | bedrock-runtime (in-Region only) |
| Input / output | Speech and text / speech and text |
| Context / max output | 256K / 64K tokens |
| Regions | us-east-1, us-west-2, eu-north-1, ap-northeast-1 |
| Service tier | Standard only (no Priority, Flex or Reserved) |
| Default quota | 20 concurrent bidirectional sessions per account per Region |
| Lifecycle | Active; end of life no sooner than 5 October 2027 |
No geo or global cross-Region inference is offered, so a European deployment means Stockholm.
Speech understanding and speech generation happen in one model. Around that core, Amazon lists:
Bedrock features you might expect are missing: Guardrails, Knowledge Bases, Agents, Flows, prompt management and token counting are all unsupported. Safety filtering and retrieval have to sit in your own code.
Amazon gives two figures: 256K tokens in Bedrock's specification for this model, and "up to 1M tokens" in its broader Nova product material. These cannot both describe the same deployed limit. Design long sessions around 256K, and if you need more, open a session that deliberately exceeds it in a test account and see what the service returns.
Per-minute or per-token rates are published on Amazon's Bedrock pricing page, which we did not verify for this profile, so no figure is quoted here. Check that page and your own account before building a cost model.
Twenty concurrent sessions is a small default for a contact centre. Request a quota increase early, then load-test at your expected peak with recorded calls. Track time to first audio, how often an async tool result arrives after the caller has moved on, and how well code-switching holds up for your bilingual customers.
For comparison, put the same scripts through Gemini 3.8 Live or GPT-Realtime-2.1. Amazon's text models, such as Nova 2 Lite, are on the Amazon provider page.
Compare the model on the work you need to complete.
Nova 2.5 Sonic is listed for Audio-to-audio, Text → Speech, Speech → Text. It accepts speech, text, audio and produces speech, text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 256K context and 64K maximum output. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
Use the model identifier and account requirements for the selected API route. Confirm its request limits and output format with one complete task before connecting a larger workflow.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Alibaba
Qwen-Audio-3.1-Realtime Flash is Alibaba's lighter full-duplex speech model with audio and text in and out, function calling, web search and a 262K context.
ElevenLabs
Eleven Realtime v1 Max is ElevenLabs' reasoning speech-to-speech agent for support and multi-step tool workflows, OpenAI Realtime-compatible.
ElevenLabs
Eleven Realtime v1 Mini is ElevenLabs' fast, low-cost speech-to-speech voice agent stack, with an OpenAI Realtime-compatible API and 90+ languages.
xAI
Grok Voice 3.0 is xAI's speech-to-speech model for real-time voice agents, replacing Grok Voice 2.0 at the same $0.08-per-minute price.
Alibaba
Qwen-Audio-3.1-Realtime-Plus supports duplex speech conversation, voice cloning, tools and a 262K-token context through its realtime interface.
Gemini 3.8 Live is Google's model for interactive audio conversations with text, image, audio and video input. Its Live API response modality is audio; text is available through transcription.