Amazon Nova 2.5 Sonic: four regions, seven languages, 256K context

Amazon Nova 2.5 Sonic is a speech model for live voice applications: AWS announced it on October 5, 2026 for US East (N. Virginia), US West (Oregon), Europe (Stockholm) and Asia Pacific (Tokyo), with seven languages and a 256K context window. It is built for agents that listen and answer in a session. A finished voiceover for a short video is a different job, and on Sume that is a text-to-speech file job.

Every Nova fact here comes from the AWS announcement read on 2026-10-13; AWS says pricing is the same as Nova 2 Sonic and does not restate it in the announcement.

Does the region list matter for a video team?

Only if you run a live agent: a caller in Tokyo wants the session in Tokyo. A rendered voiceover has no caller, so region is not a quality choice. Sume's TTS Router lets you pick the engine with a required model field (for example sonic-3.6), not a cloud region.

What should a short-video team use instead?

Generate the line as a job and fetch the file. POST /v1/tts-router/generate takes a model, one transcript of 1 to 20000 characters and a voice, then you poll GET /v1/jobs/:id/status and read the audio from /result. Sonic rows list at 38 micro-dollars per character before the 1.25 margin, which is about $0.0475 per 1,000 characters before cent rounding, matching the TTS 1.0 catalog row.