On September 10, 2026, OpenAI released GPT-Live-1 in the API. We are excited to share that Savvy Agents has added GPT-Live-1 support to Ira, our AI dental receptionist.
We did not evaluate the release with a scripted demo alone. We put it through controlled internal phone tests built around the moments that make dental conversations difficult: verifying a patient, finding an appointment, changing a decision, handling interruptions, sending a follow-up text, and ending the call correctly. The results reinforced a conviction we have held from the beginning: a good dental voice agent needs both a natural conversation and dependable action behind it.
What GPT-Live-1 changes about real-time voice
What it is: GPT-Live-1 is OpenAI's new real-time voice model for applications that need a spoken conversation to continue while useful work happens in the background. Its full-duplex behavior lets the system listen and speak at the same time, which is closer to the rhythm of an ordinary phone call than a rigid question-and-answer exchange. OpenAI describes the release in its GPT-Live-1 announcement.
That capability matters because people do not speak in clean turns. A patient may add a qualification, correct a detail, or ask a second question before the receptionist has finished responding. They may pause to check a calendar or decide that the afternoon works better than the morning. A useful voice system has to hear those changes and respond to the latest approved intent.
The model is only one part of the experience. GPT-Live-1 does not know a practice's hours, provider rules, appointment types, identity policy, or scheduling state by itself. Savvy Agents connects the voice layer to those workflows and gives Ira a clear boundary between what can be said, what must be confirmed, and what can change in the dental system.
Our approach to a dental conversation
Dental calls are short, but they carry consequential details. A patient's name and date of birth determine who the practice is speaking with. An appointment time affects the patient, the provider, and the operatory. A cancellation changes the schedule. The conversation therefore has to be warm without becoming casual about the details that matter.
At Savvy Agents, we design Ira around a few practical principles:
Ask for the next useful detail. Ira should move the call forward one decision at a time. If the patient has already answered one part of a request, the next question should fill the remaining gap instead of starting the intake over.
Keep information the caller has already supplied. Repeating a name or appointment preference can make a conversation feel mechanical. Ira should use a detail after it has been clearly provided and verified.
Confirm consequential actions. Natural speech does not mean guessing. Before booking, moving, or canceling an appointment, Ira needs the patient's approval of the specific action and time.
Keep internal work out of the spoken exchange. Patients need to know what is happening for them. They do not need a narration of every internal lookup, response, or software step.
Fail safely when identity or intent is unclear. A polite handoff or request for clarification is better than an appointment change based on an assumption.
These principles apply regardless of the selected voice model. GPT-Live-1 gives us a stronger foundation for turn-taking and listening, while Savvy's workflow layer keeps the conversation tied to practice rules.
How we tested GPT-Live-1 support in Ira
What we tested: We ran internal phone calls through an isolated dental practice sandbox. The calls exercised patient verification, appointment booking, rescheduling, cancellation, interruption recovery, SMS follow-up, and automatic call closing. We used separate scenarios because each action has its own failure modes, and a successful booking does not prove that a cancellation or a closing signal works the same way.
For every action that changed scheduling data, we compared what Ira said with what the connected dental system recorded. That showed whether the spoken confirmation matched the appointment actually created or changed. We also checked approved text messages and automatic call closing.
Some early scenarios exposed work that needed refinement. One completed an appointment action but did not close automatically; another reached the right scheduling result while the follow-up text path needed recovery. We treated those outcomes as findings, refined the relevant state and action handling, and repeated focused scenarios. Latest local verification confirmed the intended SMS delivery and automatic closing behavior in fresh calls.
That is our standard for dental voice AI. A transcript that sounds good is not enough, and a technically correct record is not enough if the patient cannot follow the conversation. We inspect both sides.
What we learned
Conversation quality starts with pacing
In one focused test, we kept the Marin voice the same while refining prompt pacing and handoff guidance. Ira asked for the caller's name, waited, asked for the date of birth, waited again, and then gathered the appointment type and preferred day. That rhythm changes how the call feels. The more natural exchange we heard in follow-up testing reflected those conversation refinements alongside GPT-Live-1's capabilities, rather than a model-only measurement.
We also saw the value of confirming information that the caller had already provided naturally. Ira could acknowledge the name and date of birth without forcing the caller through an unnecessary spelling exercise. When an important detail was unclear, the conversation could still fall back to a more explicit confirmation. The goal is not to remove safeguards. It is to apply them at the moment they help.
A changed decision must reach the schedule
People change their minds during phone calls. In a scheduling scenario, the caller moved from a morning request to an afternoon appointment. Ira recognized the correction, confirmed the new choice, and the connected dental system recorded the approved afternoon slot. The test mattered because it joined conversational understanding to a real scheduling action.
Voice quality and workflow accuracy meet here. A system can sound attentive while retaining an earlier preference, or make the right database change while explaining it poorly. We look for both outcomes together.
Closing is part of the experience
The last seconds of a call deserve the same care as the greeting. A patient should not have to wonder whether the appointment was saved, whether a text was sent, or whether the call is still active after saying goodbye. Our tests therefore included automatic closing and follow-up delivery as explicit acceptance points.
In fresh verification, an approved text message was delivered through the configured provider and the application ended the call automatically after the conversation was complete. We also tested interruption and recovery paths separately, because a caller speaking during a closing phrase can change the expected state. This approach helps us distinguish ordinary completion from recovery behavior instead of treating every hang-up as the same event.
What GPT-Live-1 support means for practices
GPT-Live-1 support is deployed in Savvy Agents as a selectable option for configured phone numbers. That gives us a practical way to apply the model where a practice's workflow, scheduling configuration, and safeguards are ready for it. It also lets us keep evaluating the experience with the same discipline we use for every voice change.
Best for: Practices that want a more natural phone experience while keeping patient verification, scheduling rules, appointment confirmations, and team handoffs connected to the systems they already use.
Limitation: A new voice model does not remove the need for careful practice configuration or workflow testing. Availability depends on the number and setup being evaluated, and internal test results should not be read as a guarantee that every future conversation will behave perfectly.
Our work brings together spoken interaction, practice knowledge, and the action layer that makes an answer useful. When a patient asks for an appointment, Ira must understand the request, verify who is calling, respect available options, obtain approval, and leave an accurate record.
You can learn more about Ira for dental practices and how Savvy connects voice conversations to day-to-day front desk work.
Going beyond the phone call
Ira is one part of a broader set of Savvy Agents built around the work dental teams do every day. Sia supports clinical workflows, helping practices carry important work forward after the patient conversation. Milo focuses on insurance-related workflows, and Novi helps with patient reminders and follow-up.
Those agents do not all use GPT-Live-1, and this announcement is specifically about GPT-Live-1 support in Ira. The broader point is that a practice's AI experience should extend beyond a single call. Voice, clinical work, insurance, and reminders each have different requirements, so each agent needs the right model and controls for its job.
For practices using Open Dental, our Open Dental AI receptionist integration page explains how appointment conversations can connect with that environment. The exact experience depends on the practice's configuration and the workflow being used.
What comes next
GPT-Live-1 is new, and there is already a lot to explore. We are excited by how quickly the technology is improving, while staying grounded in the details that determine whether a dental practice can trust a voice interaction.
We will keep running focused calls, reviewing the spoken experience alongside the resulting records, and refining how Ira handles interruptions, confirmations, handoffs, and closing. The best evidence will continue to come from real workflow checks, not from a polished demo that avoids the hard moments.
See Ira in action
Want to explore what Ira could do for your front desk? Book a demo with our team to discuss your practice's calls, scheduling needs, and whether GPT-Live-1 support is a fit for your setup.
