Bedrock adds GLM 5.3 for approved customers and ships Nova 2.5 Sonic for voice agents

AWS announced two model additions to Amazon Bedrock on October 5, 2026. The first is GLM 5.3 from Z.ai, the open-weight model that Anthropic's red team analyzed in September for its exploit-writing ability. The second is Amazon Nova 2.5 Sonic, a new version of AWS's own speech-to-speech model for voice agents. One comes with a gate and a price list; the other comes with neither a benchmark nor a new price.

GLM 5.3: the numbers on the card

The Bedrock model card lists a 1 million token context window and up to 128K output tokens. Reasoning is always on, and you pick an effort level to trade latency and token use against task performance. The model supports function calling, structured JSON output, and streaming, including streamed reasoning content. Prompt caching is both implicit and explicit, with a minimum of 1,024 tokens per cache checkpoint and retention of at least 30 minutes.

On the Bedrock pricing page, GLM 5.3 through the global cross-region profile costs $1.68 per million input tokens and $5.28 per million output tokens. Cache reads are $0.312 per million tokens and a 30 minute cache write is $2.10 per million. The announcement describes the model as a mixture of experts with 753 billion total parameters and roughly 40 billion active per token; the model card says 744 billion total, and the two AWS pages do not agree on that figure.

GLM 5.3: the gate

The model card states that access is limited to eligible customers and tells you to contact your AWS account team. The announcement is addressed to enterprise customers. There is also no in-region inference: requests must use the US geographic inference profile, us.zai.glm-5.3, or the global profile, global.zai.glm-5.3, and the global profile can be reached from regions across North America, Europe, Asia Pacific, the Middle East, Africa, and South America.

If you do not have an account team to ask, you cannot use GLM 5.3 on Bedrock today. For most independent developers, the listed price is therefore a reference point for comparing other hosts of the same open weights rather than a price you can pay.

Nova 2.5 Sonic: what AWS says changed

AWS says Nova 2.5 Sonic improves reasoning, instruction following, and tool-calling accuracy, and lowers latency. The announcement lists seven languages, controllable turn-taking, voice and text in the same session, and asynchronous tool calling. It is available in US East (N. Virginia), US West (Oregon), Europe (Stockholm), and Asia Pacific (Tokyo), and AWS says it keeps the same price as Nova 2 Sonic.

What the announcement does not contain is a benchmark. There is no number for latency, word error rate, or tool-calling accuracy, so the improvement over Nova 2 Sonic cannot be measured from AWS's own text. The claims are the vendor's description of its own model.

Strands Bidi Agents are generally available

On the same day, the open source Strands Agents project declared its Bidi Agents, the bidirectional voice agents, generally available, which means the API is now stable under its versioning policy. They support Nova 2.5 Sonic, OpenAI's GPT-Realtime-2.1, and Google's Gemini 3.8 Live. According to Strands, a Nova 2.5 Sonic session is limited to 8 minutes, and Bidi Agents extend it by reconnecting in the background automatically. That limit is the one concrete constraint on the new model that either source states.