PolyAI's Dialog-RSN-1 Hits SageMaker JumpStart at $45 Per Host Hour
AWS added PolyAI's audio-native customer-service model Dialog-RSN-1 to SageMaker JumpStart on November 19, 2026, promising one-click deployment — with text-to-speech, telephony and 24/7 billing left to the customer.

Why it matters
- AWS added PolyAI Dialog-RSN-1 to the SageMaker JumpStart catalog on November 19, 2026, three months after the model's July 30 launch.
- Real-time inference costs $45 per host hour on ml.p4d.24xlarge and $110 on ml.p5.48xlarge, billed from service start to deletion even when idle.
- PolyAI reports median first-token latency below 300 ms and about 390 ms for a typical full caller exchange in production.
- The model accepts PCM16 mono audio at 16 or 24 kHz only, has a 32,000-token context window, and is English-only.
- Connections cap at 30 minutes with a five-minute idle timeout, and state does not survive reconnects.
AWS added PolyAI's Dialog-RSN-1 to the Amazon SageMaker JumpStart catalog on November 19, 2026, giving the audio-native customer-service model one-click deployment, evaluation tools and four optimization profiles — at $45 per host hour on an ml.p4d.24xlarge and $110 on an ml.p5.48xlarge. The endpoint runs inside a customer's own AWS account, which means call audio and transcripts stay on infrastructure the customer controls. For enterprises weighing data residency and audit requirements, that detail is the reason this release matters more than the convenience of one-click deployment itself.
This is the third step in a sequence, not a fresh launch. PolyAI introduced Dialog-RSN-1 on July 30, 2026. On October 6, 2026, the company announced SageMaker endpoint deployment through AWS Marketplace. JumpStart now adds discovery, evaluation and preconfigured deployment on top — though customers still subscribe through AWS Marketplace. The pattern is familiar: AWS keeps shrinking the distance between "found a model" and "running a governed endpoint."
What does JumpStart actually change?
The workflow is now a catalog operation. From the SageMaker AI console's Model Hub, a customer selects Dialog-RSN-1 and chooses Deploy. JumpStart creates a real-time inference endpoint on a preconfigured serving stack. That removes manual container selection and the need to benchmark dozens of infrastructure combinations by hand.
Teams that prefer infrastructure-as-code can do the same thing with a few lines of SageMaker Python SDK code and fold deployment into their CI/CD pipelines. The takeaway: JumpStart shortens the path from evaluation to a governed AWS endpoint, but it doesn't absorb the whole production stack.
How do the four optimization profiles work?
Before deployment, the model card displays P50 latency, time to first token and throughput. Customers then pick one of four profiles:
- Latency — minimizes time to first token
- Throughput — maximizes tokens per second
- Cost — minimizes infrastructure cost per request
- Balanced — general-purpose use
The choice ties the serving configuration to call volume, responsiveness and budget, rather than 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 kinds of calls it actually expects. Reviewing the overview and samples is free; streaming your own audio through the model requires an endpoint.
Why is audio-native different from a speech-to-text pipeline?
Dialog-RSN-1 processes caller audio directly and combines four jobs in one model: turn-taking, speech recognition, response generation and function calling. PolyAI's argument is that direct audio preserves tone, hesitation and timing — the cues a cascaded system loses when speech recognition flattens a call into a transcript.
The turn-taking mechanism is specific. A voice-activity detector triggers each probe, and the model's first output token records EMPTY, ONGOING or COMPLETE, placing the final decision about whether the caller has finished speaking with the model rather than a fixed rule. In plain terms: the model listens to how people actually talk, and it decides when to jump in.
The output path, however, is separate. Dialog-RSN-1 returns streamed reply text or a function call, with the caller transcript arriving separately. It does not synthesize speech. Every deployment must therefore add its own text-to-speech layer — a design that preserves the customer's choice of voice but pushes another component outside the model endpoint.
What latency numbers should buyers trust?
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. That distinction matters. Model latency alone doesn't describe the delay in a complete voice deployment, which also includes speech synthesis, telephony and external tool calls. Anyone budgeting for conversational responsiveness needs to measure the full chain, not the model card.
What are the integration limits?
The endpoint lives in the customer's virtual private cloud, uses its AWS Identity and Access Management roles, and reports through Amazon CloudWatch. It 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. Applications can also invoke that stream directly.
The hard constraints are worth listing plainly:
- Audio input: base64-encoded PCM16 mono at 16 or 24 kHz — no G.711, no Opus, no stereo
- Context window: 32,000 tokens
- Languages: English only
- Connections: 30-minute hard cap, five-minute idle timeout
- State does not carry across reconnects, so applications must restore relevant context when establishing a new session
Customers retain responsibility for telephony, orchestration, business logic and tool execution. PolyAI's integration guidance also calls for validating dates, reference numbers and amounts before acting on tool calls. The bottom line: the model handles the conversation, and everything around the conversation stays your job.
How does billing work — and when does it stop?
Real-time inference costs $45 per host hour on ml.p4d.24xlarge and $110 per host hour on ml.p5.48xlarge. A host hour is one hour of runtime for the compute instance, and additional AWS infrastructure charges may apply. The critical detail: billing starts when the endpoint enters service and continues until deletion — even when no client is connected. An idle endpoint accrues charges around the clock. AWS's own deployment examples delete the endpoint, endpoint configuration and model during cleanup, which reads as a quiet warning.
What's the actual business case?
JumpStart's real contribution here is a shorter operational path from model evaluation to a governed AWS endpoint. The remaining production work sits around the model: telephony, speech synthesis, tool validation, session recovery, and control of an endpoint that bills continuously whether or not anyone is calling. The decision for any buyer therefore comes down to a measurable comparison — call-handling results from the customer's own audio, weighed against those continuing engineering and hosting costs.
Source: unite.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
823 articles
Related articles
- AWS Launches SageMaker JumpStart Leaderboard Ranking 70+ Models
- OpenAI Opens ChatGPT Enterprise Sales Through AWS Marketplace
- AWS Turns Step Functions Into Agent Tools as GPT-6.1 Terra Hits Bedrock
- AWS AgentCore’s Strands upgrade can change omitted agent settings
- OpenAI Brings GPT-5.5, Codex, and Managed Agents to AWS Bedrock