OpenAI's GPT-Live-1 and Google's Gemini 3.8 Live shipped five days apart and cut real-time voice to cents a minute. For a Canadian clinic or small business, the model was never the expensive part — here's how to decide whether to build on the new APIs or buy a managed AI receptionist.
The short answer: the models got cheap this week, not the receptionist. On 10 September 2026 OpenAI shipped GPT-Live-1, a full-duplex voice layer at US$0.05 a minute; five days later Google answered with Gemini 3.8 Live. Voice is now close to a commodity input, but a working AI receptionist is still the backend model, the integrations, the call-flow design, the testing, and, for a Canadian clinic, the data-residency and PHIPA/PIPEDA work around it. Buying remains the sounder call for most Ontario clinics; building on the new APIs is now viable for a technical founder running a non-health business with low-stakes calls. This post is the framework for telling which one you are.
On 10 September 2026, OpenAI introduced GPT-Live-1 in its API: a voice model built for full-duplex conversation, meaning it can listen and speak at once instead of waiting for a caller to finish talking. OpenAI says the model handles interruptions naturally, delegates reasoning and tool calls to a backend model of your choosing, and is built for telephony: it "enables deployment of full-duplex voice agents for phone calls, from restaurant reservations to customer support." Twelve new voices shipped with it, and OpenAI reports a 30-percentage-point improvement over GPT-Realtime-2.1 on Full Duplex Bench. Yelp's CTO, Alex Levy, said GPT-Live-1 is delivering "meaningful improvements in call handling rates," and Speak reported cutting thinking-pause interruptions by almost 80% versus older turn-based systems. (OpenAI, 10 September 2026)
Five days later, Google answered with Gemini 3.8 Live and a companion Extended Thinking variant, reporting 82.6 on Artificial Analysis's Speech to Speech Quality Index and 68.6% on τ-Voice, a benchmark for agentic task completion in conversation. It detects and switches between 97 languages mid-conversation and runs tool calls in the background without interrupting speech; generated audio carries a SynthID watermark. It's rolling out in the Gemini API and AI Studio, with Gemini Enterprise in private preview. (Google, 15 September 2026)
The headline number is US$0.05 per minute, and it's tempting to read that as the whole AI voice agent cost per minute. It isn't. OpenAI's own wording is specific: "GPT‑Live‑1 is available in the API today at $0.05 per minute for the front-end voice layer. Pair it with the backend model and agent harness that fit your product…" That price covers listening and speaking. The reasoning, and any tool calls such as checking a calendar, come from a separate backend model billed separately.
A worked example: at the Bank of Canada's 16 September 2026 rate of 1.3947, US$0.05 is roughly CA$0.07 a minute. A clinic doing 1,000 call-minutes a month would pay 1,000 × US$0.05 = US$50, about CA$70, for the voice layer alone, before backend-model token charges and telephony. Twilio's Canada voice pricing lists inbound local calls at $0.0085 a minute plus $1.15 a month for a local number. This is illustrative arithmetic, not a quote. (Bank of Canada, Twilio Canada voice pricing)
Orchestration platforms that bundle the voice layer, backend model and tooling price the assembled pipeline as one per-minute rate: Retell AI's public pricing lists pay-as-you-go voice agents at US$0.07–$0.31 a minute depending on the LLM and voice chosen, with telephony on top. (Retell AI pricing)
A per-minute price for a voice model is not a receptionist. Between "the model can hold a phone conversation" and "the phone rings and gets handled correctly" sits a list of work that doesn't show up on OpenAI's or Google's pricing page:
None of this is exotic engineering, but it's real project time, whether you build on GPT-Live-1 or on Gemini 3.8 Live. See our guide to what a custom AI agent costs to build, or our Pilot-to-Production Deployment service for what the work looks like end to end.
Here's the catch for a business operating under Canadian data residency expectations. OpenAI's data-controls documentation is specific: "GPT-Live supports data residency in the United States and Europe. Delegated backend models and tools have their own data controls." OpenAI's Canada region is storage-only, with no in-region processing, and GPT-Live and Realtime endpoints aren't among the services listed for it. Sessions are eligible for zero data retention, and session storage is off by default (if you switch it on, a completed recording is kept for 30 days) — but that doesn't change where the audio is processed while the call happens. (OpenAI data controls documentation, checked 17 September 2026)
For Gemini 3.8 Live, we did not find official confirmation that Google can process the audio in a Canadian region. If that matters, and for a clinic it does, get it confirmed in writing from Google first.
The price tells you what the voice layer costs, not where your callers' voices are processed. For an Ontario clinic booking patients by phone, that's not a footnote — see our note on responsible AI in Canada for how we think about Canadian data residency at the architecture stage.
Two Canadian privacy regimes touch this decision, and neither one gives a simple yes-or-no answer on where data has to live.
PIPEDA doesn't ban cross-border processing. The Office of the Privacy Commissioner of Canada's guidance is direct: "PIPEDA does not prohibit organizations in Canada from transferring personal information to an organization in another jurisdiction for processing." But accountability stays put: "The transferring organization is accountable for the information in the hands of the organization to which it has been transferred." Customers must be told their information may cross the border and could be accessed by that country's courts or authorities. (Office of the Privacy Commissioner of Canada)
Ontario's PHIPA adds a layer for health information custodians — a dental or physiotherapy clinic answering the phone about a patient's care. There's no blanket rule that health information must stay in Canada; PHIPA requires the custodian take reasonable steps to safeguard it (section 12) and remain responsible for vendors handling it on the custodian's behalf. A caller explaining why they need an appointment is disclosing health information the moment they say it. The Information and Privacy Commissioner of Ontario's guidance, "AI Scribes: Key Considerations for the Health Sector" (28 January 2026), is the closest regulator guidance to an AI system that records patient conversations, and is worth reading directly. (IPC of Ontario)
None of this is legal advice; talk to a privacy lawyer about your specific obligations. If federal AI legislation is relevant to your business, see our separate note on Bill C-36. Keeping processing and storage in Canada isn't the only lawful route, but it removes the cross-border questions entirely — one reason we built Maya that way.
Laid out side by side, the three paths differ less in what's technically possible than in who owns the work of getting there.
| Factor | DIY on the raw API | Custom build with a partner | Managed receptionist |
|---|---|---|---|
| Setup time | Weeks to months, depends on in-house time | Typically weeks, scoped up front | Live in days |
| Who owns integration & upkeep | You | Partner builds it; you or the partner maintain it | The vendor |
| Data residency | Whatever the model vendor offers — confirm in writing | Chosen deliberately at the architecture stage | Set by the vendor's existing architecture |
| PHIPA / PIPEDA paperwork | Yours to work out | Built into the scoping phase | Handled by the vendor's existing setup |
| Ongoing engineering | Required indefinitely | Needed for changes, less for day-to-day | Minimal — the subscription covers upkeep |
| Cost shape | Variable, per-minute plus your time | Fixed project fee, scoped up front | Fixed monthly subscription |
| Best for | Technical founder, non-health business, low stakes | Regulated or complex workflow, wants ownership | Clinic or SMB that wants it live without building a team |
Building a custom AI voice agent on GPT-Live-1 or Gemini 3.8 Live is reasonable for a non-health business with an engineer to own the integration, and calls where a wrong answer costs little more than a redo — a quote line, an order-status line, an internal helpdesk.
What building doesn't remove is the list above: designing the call flow, connecting it to your booking system, testing it against real callers, watching it once live. If that's more time than you have, price out a scoped custom build first — see our AI agent cost guide for Canadian numbers.
Buying a managed receptionist is the sounder default for a Canadian clinic. Health information comes up on many calls, which puts you inside PHIPA's custodian duties from the first ring. A vendor that has already made the data-residency and compliance decisions removes that research project from your to-do list.
It's also right for any small business without engineering time for ongoing upkeep. Mihron AI's Maya is an AI phone answering service for exactly that: an AI voice receptionist for dental clinics and other Ontario SMBs, bilingual in English and French (see our note on bilingual AI receptionists), purpose-built for dental clinics including Canadian Dental Care Plan questions, hosted in a Canadian Supabase region with row-level security per clinic. Plans start at CA$299 a month. Our buyer's checklist covers what to ask before you sign any managed receptionist.
To be direct: Maya does not run on GPT-Live-1 or Gemini 3.8 Live. Its stack today is Retell AI for orchestration, Anthropic's Claude API for reasoning, Deepgram Nova-3 for speech-to-text and Cartesia Sonic for text-to-speech, over Twilio and Vonage telephony, with data in a Canadian Supabase region.
We're not ignoring what shipped this week. Before routing any call through a new voice model, we'd benchmark it against our current stack using test scripts and consented call recordings. We wouldn't route an Ontario clinic's calls through an endpoint that can't process audio in Canada, or under terms a health custodian can't accept — which rules out GPT-Live-1 for us in that use case today, whatever the voice quality. If either vendor extends Canadian processing, that changes the calculus. This is the first question our AI Readiness Sprint (CA$5,000–9,500) answers, feeding into a fixed-fee, scoped-up-front AI Agent & Workflow Build.
If you're deciding between building and buying, start with the honest question: what happens on your calls, who owns the integration a year from now, and where does the audio need to live. A scoped AI Readiness Sprint answers that with your own numbers. If a managed receptionist fits better, hear Maya on our live demo or call 437-367-8009. For build vs buy at enterprise scale, see our related post on build vs buy for enterprise GenAI, and for the guardrails question next door, see AI agent guardrails.
Hear Maya answer a call, or book a scoped readiness sprint before you write a line of integration code.