AWS’s November 19, 2026 JumpStart walkthrough shows customer-service teams how to deploy PolyAI Dialog-RSN-1 in their own AWS account. At the listed p4d software rate, eight billed host-hours cost $360 for the model alone. A voice pilot also needs AWS infrastructure, separate speech synthesis and application integration.
This is a deployment and purchasing decision, not a new-model launch: PolyAI introduced Dialog-RSN-1 on July 30 and announced SageMaker availability on October 6.
What does an eight-hour pilot cost?
The Marketplace usage table lists model-software charges of US$45 per host-hour for ml.p4d.24xlarge real-time inference and $110 for ml.p5.48xlarge real-time inference. A separate ml.g6e.4xlarge batch option is $120 per host-hour. These prices do not establish how many simultaneous calls each instance can handle.
For an illustrative pilot with one p4d host, using the public list price checked on November 19 and assuming no discounts or credits:
Eight billed hours: 1 × 8 × $45 = $360 in model-software charges.
The same host running for 24 billed hours: 1 × 24 × $45 = $1,080 in model-software charges.
Those are hours the host runs, not hours callers spend talking. AWS’s Marketplace billing rules separate software from infrastructure and prorate hourly software usage to the minute. Neither example is a complete pilot quote or an automatic spending cap.
For each instance type, calculate billed host-hours × (software hourly rate + the applicable SageMaker inference infrastructure hourly rate). Add those amounts, plus speech synthesis, telephony, any storage, logging and data-transfer charges, integration work and applicable taxes. The infrastructure and voice-provider amounts remain unpriced here.
Closing a notebook or disconnecting a caller does not remove the endpoint. Assign someone to delete the endpoint after the pilot; deleting its configuration alone does not stop it. Nor can this Marketplace GPU deployment use SageMaker Serverless Inference, which excludes GPUs and Marketplace model packages.
A cost per resolved call needs workload-specific capacity and successful-outcome data, not just an hourly tariff. That distinction also underlies budgeting Bedrock coding agents beyond model prices, although Bedrock’s rates do not apply to this deployment.
What must you add to make a phone agent?
Dialog-RSN-1 processes incoming audio but returns text or a request to use a business tool. PolyAI’s integration guide tells developers to supply the text-to-speech engine that turns replies into a speaking voice. The application also handles playback interruptions and executes tool requests, validating dates, amounts and reference numbers before acting.
Separating the voice engine gives the team a choice of speaking voice, while adding another cost and integration decision. A working text response is therefore not proof of a working phone service or its end-to-end response time.
For existing WebSocket-based Realtime clients, PolyAI supplies a bridge to SageMaker’s signed HTTP/2 stream. A client can instead call that AWS streaming interface directly. The bridge connects protocols; it does not supply speech output or the surrounding phone application.
Decide which business actions the application may execute before connecting it to customer systems. For a related design choice on another platform, see choosing web search or business tools for a Qwen voice assistant.
Which limits could rule it out?
The current Marketplace listing specifies English support. A requirement for other languages needs resolving before buying an endpoint, rather than relying on a roadmap.
PolyAI’s sample notebook documents two further integration constraints:
Audio format: PCM16 mono at 16 or 24 kHz. G.711, Opus and stereo need conversion before reaching the model.
Conversation memory: 32,000 tokens shared by instructions, tools, knowledge, history and audio. The sample describes replacing older audio with transcripts, but text-heavy conversations can still exceed the limit.
There is also a 30-minute limit per SageMaker bidirectional stream. PolyAI’s notebook documents a five-minute idle timeout and advises restoring necessary context when reconnecting. A longer call therefore needs application-managed reconnection; a fresh stream is not a guarantee that the conversation continues with its history intact.
Hosting the model in your AWS account does not settle where every part of a call goes. AWS requires network isolation for Marketplace model containers, but your chosen speech, telephony and logging services have their own data flows. Review those components separately if location or retention requirements drive the decision.
What should you confirm before deployment?
Before starting the paid endpoint, settle three items:
Access: confirm the Marketplace subscription, regional model-package identifier, SageMaker permissions and quota for at least one selected endpoint instance in the deployment prerequisites. Region-by-region availability and your account’s GPU capacity are not established by this guide.
Budget and ownership: obtain the regional infrastructure quote, choose a speech engine, agree the pilot duration and name the person responsible for endpoint cleanup.
Commercial terms: review the Marketplace offer and its linked model EULA separately from the sample repository’s Apache-2.0 code licence. The code licence is not the model’s licence.
The documented fit is an English-language use case with a reason to control the voice and business integrations, and a team able to operate the AWS deployment. If the requirement is a ready-made phone service, multilingual coverage or a proven low cost per call, resolve those gaps before committing to host-hours.
