Innovation & Tech

PolyAI's Dialog-RSN-1 Lands in SageMaker JumpStart at $45 an Hour

AWS added PolyAI's audio-native Dialog-RSN-1 to SageMaker JumpStart on November 19, 2026, with sub-300ms first-token latency and $45-per-hour hosting — while customers still build telephony and speech synthesis.

By Daniel Okafor

4 min read

AWS Adds PolyAI Dialog-RSN-1 to SageMaker JumpStart Catalog
AWS Adds PolyAI Dialog-RSN-1 to SageMaker JumpStart CatalogAI-generated

What's News

  • AWS added PolyAI Dialog-RSN-1 to SageMaker JumpStart on November 19, 2026 with four optimization profiles.
  • Real-time inference costs $45 per host hour on ml.p4d.24xlarge and $110 on ml.p5.48xlarge, billed while idle.
  • Median first-token response is below 300 milliseconds; a typical full exchange takes about 390 milliseconds.
  • The model returns text or function calls only, so deployments must add their own text-to-speech layer.
  • Connections cap at 30 minutes with a 32,000-token English-only context window.

AWS added PolyAI's audio-native customer-service model Dialog-RSN-1 to the Amazon SageMaker JumpStart catalog on November 19, 2026, with four optimization profiles and pricing that starts at $45 per host hour on ml.p4d.24xlarge and rises to $110 per host hour on ml.p5.48xlarge.

The catalog entry shortens the path from evaluation to a governed AWS endpoint. It does not deliver a finished voice assistant. Dialog-RSN-1 returns only text or function calls, so every deployment must add its own text-to-speech layer, and billing starts when the endpoint enters service and continues until deletion, even with no client connected.

What does JumpStart actually add?

This is the third step in a sequence. PolyAI introduced Dialog-RSN-1 on July 30, 2026 and announced SageMaker endpoint deployment through AWS Marketplace on October 6, 2026. JumpStart now layers discovery, evaluation tools and one-click deployment on top, while customers still subscribe through AWS Marketplace.

From the SageMaker AI console's Model Hub, a customer selects Dialog-RSN-1 and chooses Deploy. JumpStart builds a real-time inference endpoint on a preconfigured serving stack, removing manual container selection and the need to benchmark dozens of infrastructure combinations by hand. Teams can drive the same deployment with a few lines of SageMaker Python SDK code and fold it into CI/CD pipelines.

The model card shows P50 latency, time to first token and throughput before deployment. Four profiles — latency, throughput, cost and balanced general-purpose — map the serving configuration to call volume, responsiveness and spending targets instead of applying one default to every workload.

Evaluation lives in the same catalog entry. PolyAI's sample notebook streams recorded caller audio and prints the model's turn decisions, replies and transcripts, so a team can inspect behavior on the calls it expects to handle. Streaming a customer's own audio requires an endpoint.

How does the model handle speech?

Dialog-RSN-1 processes caller audio directly and combines turn-taking, speech recognition, response generation and function calling in one model. PolyAI says direct audio preserves tone, hesitation and timing that a cascaded system can lose when speech recognition reduces a call to a transcript. A voice-activity detector triggers each probe, and the model's first output token records EMPTY, ONGOING or COMPLETE, placing the final turn-taking decision with the model itself.

The output path is separate. The model returns streamed reply text or a function call, with the caller transcript arriving separately. It does not synthesize speech. A deployment must supply text-to-speech — a component outside the model endpoint that preserves the customer's choice of voice.

PolyAI's Marketplace listing reports a median first-token response below 300 milliseconds and about 390 milliseconds for a typical full caller exchange in production. The gap matters: model latency alone does not describe the delay in a complete voice deployment that includes speech synthesis, telephony and external tools.

What are the limits and the costs?

The endpoint runs inside the customer's virtual private cloud, uses its AWS Identity and Access Management roles and reports through Amazon CloudWatch. Call audio and transcripts stay within infrastructure the customer controls — a design relevant to data-residency and audit requirements.

The endpoint implements the OpenAI Realtime protocol. PolyAI's client bridge exposes a local WebSocket for Realtime clients and carries their events over SageMaker's signed bidirectional HTTP/2 stream, which applications can also invoke directly. Customers retain responsibility for telephony, orchestration, business logic and tool execution. PolyAI's integration guidance calls for validating dates, reference numbers and amounts before acting on tool calls.

The technical constraints are specific:

  • Audio input: base64-encoded PCM16 mono at 16 or 24 kHz; no G.711, Opus or stereo
  • Context window: 32,000 tokens; English only
  • Connections: 30-minute hard cap, five-minute idle timeout
  • State does not carry across reconnects; applications must restore context on a new session
  • A host hour is one hour of instance runtime, and additional AWS infrastructure charges may apply

AWS's deployment examples delete the endpoint, endpoint configuration and model during cleanup — a habit worth copying, given the idle-billing structure.

The business case now rests on results, not convenience. JumpStart removes the setup work, but the production work remains around the model: telephony, speech synthesis, tool validation, session recovery and an endpoint that accrues charges around the clock. Buyers will need call-handling results from their own audio to justify those continuing engineering and hosting costs.

Source: unite.ai

Share this article:

More from Daniel Okafor

Daniel Okafor

Show full bio

Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.

2,420 articles

Related articles

« Previous article