Getting Started With Nova 2.5 Sonic
Nova 2.5 Sonic is Amazon's speech-to-speech model for building real-time, human-like conversational AI on Amazon Bedrock, with better reasoning, lower latency and higher cost efficiency.
Nova 2.5 Sonic is Amazon's speech-to-speech model for building real-time, human-like conversational AI on Amazon Bedrock. It is designed for teams building voice agents, agentic calling, support automation and other interactive voice experiences where users expect fast responses, reliable instruction following, and accurate tool use. Sonic 2.5 is built with a newer reasoning component and closes the #1 gap customers flag in Sonic 2 on reasoning capabilities: it is smarter (AudioMC +43%, ComplexFuncBench +25%, tau3-bench +20%), faster (TTFA P50 -28%, P99 -41%), and cheaper per hour at the same per-token pricing in IAD with terser output.
What’s new
Nova 2.5 Sonic is the reasoning upgrade to Nova Sonic. Compared with Nova 2 Sonic, it is built to make voice agents:
- More reliable in real workflows: better tool calling, stronger instruction following, and better multi-step task completion.
- Faster to start speaking: lower time to first audio on short conversational turns.
- More efficient in conversation: terser answers can reduce cost per hour even when per-token pricing is unchanged.
- Easy to deploy from Sonic 2: same API surface, voices, and language/locale coverage; migration is primarily a model ID change plus prompt tuning.
The main migration consideration is prompt tuning. Nova 2.5 Sonic follows instructions more strictly and tends to answer more concisely, so prompts written for Nova 2 Sonic may need adjustment to preserve the right amount of detail and avoid unnecessary refusals.
Performance overview
Nova 2.5 Sonic brings performance improvements over Nova 2 Sonic on many dimensions, all targeting at resolving questions that customers care most about.
| Customer question | What improved | Result (vs Nova 2 Sonic) |
|---|---|---|
| Will my voice agent follow instructions better? | Audio MultiChallenge | +43% |
| Will tool calling be more reliable? | ComplexFuncBench | +25% |
| Can it complete multi-step agentic tasks? | tau3-bench, Airline and Telecom domains | +20% |
| Will users hear responses sooner? | Time to first audio, P50 | 1.50s to 1.08s, a 28% reduction |
| What about tail latency? | Time to first audio, P99 | 2.45s to 1.45s, a 41% reduction |
| Is the conversation quality better? | Human blind-vote win rate | 68.3% for male voice and 65.6% for female voice |
| Will it cost less to run? | Estimated cost per hour | Lower expected cost per hour because responses are shorter; external measurement pending |
Early testing showed the same pattern in real applications from our customers:
- A production evaluation saw tool calling improve from 73% on Sonic 2 to 90%+ on Sonic 2.5, with latency within 1.5 seconds.
- An agentic calling evaluation saw task completion improve by 18% versus Sonic 2 and reach parity with GPT Realtime in its evaluation.
- A telephone-audio evaluation of 308 identical real 8 kHz cases saw accepted answers improve from 58% to 70%, invented content drop from 58 cases to 7, and answers become about 50% shorter.
- A retail tool-calling audio A/B benchmark improved from 47.8% to 78.3%, winning every category with zero regressions.
How to start
1. Select a representative Sonic 2 workflow.
Start with a voice-agent flow that uses your real system prompt, tools, function schemas, and downstream APIs.
2. Change the model ID.
Use the Nova 2.5 Sonic model ID
amazon.nova-2-5-sonic provided for your Bedrock environment and run in an available launch region. Planned GA launch regions are us-east-1, us-west-2, eu-north-1, and ap-northeast-1.3. Re-tune the system prompt.
Review instructions that control brevity, refusal behavior, tool use, and persona. Sonic 2.5 may follow terse or restrictive instructions more literally than Sonic 2.
4. Test the flows customers actually notice.
Prioritize tool selection, parameter accuracy, multi-turn context, agentic task completion, latency feel, concise-but-complete answers, and correct handling of benign and unsafe requests.
5. Compare against your Sonic 2 baseline.
Track task completion, tool-call success, turn-level latency, user-perceived responsiveness, escalation or fallback rate, and total session cost.
Below are some example evaluation metrics you could use to understand how Nova 2.5 Sonic works with your use case, please adjust the list based on your business and technical requirements.
- Tool calls: correct function, correct parameters, no unnecessary calls.
- Instruction following: stays in persona, honors formatting and brevity requirements, preserves required detail.
- Agentic flows: completes multi-step tasks without losing state.
- Latency: first audio feels responsive on short conversational turns.
- Safety behavior: fewer benign over-refusals while still handling unsafe requests correctly.
- Cost: compare cost per completed task or cost per hour, not only per-token price.
Nova 2.5 Sonic helps teams build more reliable, intelligent voice-first applications with the same Sonic API experience. It improves the parts customers most often notice in production voice agents: instruction following, tool calling, task completion, latency, and conversational efficiency.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article