Amazon Releases Amazon Nova 2.5 Sonic Speech-to-Speech Foundation Model

Introduction

On October 5, 2026, Amazon officially expanded its frontier AI lineup with the release of Amazon Nova 2.5 Sonic. Available immediately via Amazon Bedrock, this latest iteration in the Nova Sonic lineage represents a significant leap in multimodal intelligence, specifically targeting the burgeoning market of real-time, agentic voice applications.

Unlike traditional pipelines that chain separate Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) models, Nova 2.5 Sonic is a native speech-to-speech foundation model. This architecture allows for much lower latency and a more human-like conversational flow, making it an essential tool for developers building the next generation of voice-first AI agents.

• Release Date: October 5, 2026
• Platform: Amazon Bedrock
• Primary Function: Real-time bidirectional speech-to-speech
• Model Lineage: Successor to Nova Sonic (April 2025) and Nova 2 Sonic (December 2025)

Key Features & Architecture

The core strength of Nova 2.5 Sonic lies in its unified multimodal architecture. By processing audio signals directly, the model achieves natural turn-taking, allowing it to handle user interruptions gracefully through configurable turn detection. This eliminates the awkward pauses often found in legacy voice bots.

The model supports a massive 256K-token context window, with the ability to generate up to 64K output tokens. This allows for extremely long, complex conversations that maintain coherence over extended sessions. Furthermore, it supports cross-modal inputs, meaning developers can mix text and voice within a single session, even streaming real-time text transcripts of the audio conversation.

• Context Window: 256K tokens
• Max Output: 64K tokens
• Turn-taking: Configurable bidirectional interruption handling
• Multimodality: Mixed text and voice input support
• Polyglot Capabilities: Natural multi-language support with mid-conversation code-switching

Performance & Benchmarks

Nova 2.5 Sonic demonstrates marked improvements over its predecessor, Nova 2 Sonic, particularly in reasoning, instruction following, and tool-calling accuracy. These improvements are critical for 'agentic' workflows where the model must not only talk but also act on behalf of the user.

The model was evaluated using a rigorous suite of benchmarks focusing on conversational intelligence and functional reliability. While specific MMLU or HumanEval scores for the speech-specific layer are proprietary, Amazon highlights significant gains in latency reduction and concise response generation, which are the primary metrics for real-time voice interaction. Evaluation methodology included human-in-the-loop testing for naturalness and automated stress tests for tool-calling reliability in high-latency environments.