It's 11 a.m. in Montreal, you're working on your second coffee, and OpenAI just shipped three new audio models to production at the same time: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. If you're running an AI Voice Agent right now, or you're in the middle of evaluating one, your day just got more interesting.
Let me be direct: this isn't a routine update. The performance jump between yesterday's model and today's is the biggest leap on the voice side since native SIP arrived in late April. And several of the new features target the exact frictions Quebec SMBs deal with daily — French/English handling, transfers, error recovery, Law 25 compliance.
Here's what actually changes — without the marketing fluff.
What got announced, in 30 seconds
Three models, all live this morning according to OpenAI's developer announcement:
- GPT-Realtime-2 — the centerpiece. GPT-5-class reasoning baked into the voice stack, context window jumping from 32,000 to 128,000 tokens, parallel tool calls, native preambles like "one moment, let me check that."
- GPT-Realtime-Translate — live voice translation, 70+ input languages into 13 output languages. $0.034 per minute.
- GPT-Realtime-Whisper — streaming transcription. $0.017 per minute.
Pricing for the flagship: $32 per million audio input tokens, $64 per million audio output tokens — about 20% cheaper than the gpt-4o-realtime-preview most teams have been running for the last 18 months.
The audio reasoning score (Big Bench Audio) went from 81.4% to 96.6% between the previous model and this morning's release in "high reasoning" mode. That's not a cosmetic improvement. It moves an AI Voice Agent from "follows a script" to "actually understands the situation."
OK. So what does that mean for you, the owner of an SMB in Brossard, Sherbrooke, or anywhere on the north shore?
1. Five reasoning intensity levels — the latency vs. intelligence trade-off is over
Until today, picking a voice model meant picking a compromise: a fast model that responds in 320 ms but stumbles the second a customer goes off-script, or a slower model that thinks but leaves your customer waiting two seconds before each reply.
GPT-Realtime-2 introduces five adjustable reasoning levels: minimal, low, medium, high, xhigh. Default is low, which keeps latency under the psychological 800 ms threshold. But when the agent detects a complex question — say, a customer rescheduling three appointments at once, or negotiating a credit after a complaint — it ramps up to high and uses its full capacity.
For a Quebec SMB, that means you no longer have to pick between a fast agent that sounds dumb and a smart agent that sounds slow. It's a toggle you control prompt by prompt. If you want to dig deeper into latency, we covered the topic in detail in our analysis of real AI Voice Agent latency in Quebec.
2. 4× larger context window — sessions that don't "forget"
You know that moment when a customer calls, talks for 6 minutes about order #4582, gets transferred to your secondary agent, and has to re-explain everything because the agent forgot? Right.
With 128,000 tokens of context (up from 32,000), a session can now hold the customer's full CRM history, their last 4 calls, their open order, and the live conversation — without truncating, without summarizing, without losing nuance. For a medical practice in Laval or a car dealership on the south shore, it's the difference between "I need to ask you 3 more questions" and "good morning Mrs. Tremblay, I see your file from yesterday."
It's also, for the multi-agent architectures we described in our piece on multi-agent architecture, a structural shift — agent-to-agent transfers lose less context.
3. Native preambles ("one moment, let me check") — dead air officially solved
AI Voice Agents have historically had a 2-to-5-second hole when querying a calendar, CRM, or billing system. During that time, the customer hears… nothing. It's unsettling, awkward, and it drives hangups.
GPT-Realtime-2 now handles preambles natively. When the agent decides to make a tool call, it automatically says something like "one moment, let me check your file" or "give me two seconds to confirm availability." This isn't a feature you bolt on with a timer — it's baked into the reasoning model.
ElevenLabs shipped a similar fix on April 27 (pre_tool_speech), which we covered in our guide on dead air in voice agents. The difference here is that OpenAI does it at the reasoning model level — so it's more contextual, less "scripted."
4. Spoken error recovery — the agent stops failing into silence
Before: if the CRM API went down or Google Calendar timed out, the agent would sit silent for 8-10 seconds and then say "sorry, I don't understand." The customer hung up, frustrated, and called your competitor.
Now: the agent says "I'm having trouble accessing your calendar right now — would you like me to have a team member call you back in 5 minutes?" — staying in the conversation, holding control, and offering a human option.
For a Quebec SMB, this is probably the most underrated change in this release. On a day where your Hubspot or Zoho integration hiccups, you don't lose 30% of your calls anymore — you lose maybe 3%.
5. Voice tone control — calm, empathetic, or upbeat depending on context
It's a detail, but it matters. GPT-Realtime-2 automatically adjusts its tone:
- Calm during technical troubleshooting with a frustrated customer.
- Empathetic when the customer reports a billing issue.
- Upbeat when the transaction wraps up well ("perfect, your appointment is booked for Tuesday at 10!").
You can also force it in the system prompt. For veterinary clinics, dental practices, or any business handling emotionally loaded calls (and in Quebec, we handle a lot), this is a real retention lever.
6. GPT-Realtime-Translate, separately — the bilingual fix we've waited 2 years for
Here I have to be careful: the model translates 70+ input languages into 13 output languages. French and English are in both sets, so yes, it works for Quebec. But it's not an AI Voice Agent — it's a translation model you use separately.
Typical use case: an allophone customer (Mandarin, Spanish, Haitian Creole, Arabic) calls your SMB, GPT-Realtime-Translate converts their speech into French live, your agent answers in French, and the response is translated back to the customer's language. In Montreal and Laval, where more than 30% of business clientele speaks a language other than French at home per Statistics Canada 2026 data, that's a real competitive edge.
At $0.034 per minute, on an average 4-minute call, you're talking about an extra $0.14 per allophone call. Marginal cost to stop losing those customers.
The 2 things that don't change
Don't panic, don't tear it all down. Here's what today's release does not solve:
Your Law 25 obligations stay intact. The model is more performant, but consent, recording retention, right to deletion, and the disclosure banner are still on you. If you're not sure where you stand, run the 12-question test we published last week.
Your phone integration (SIP, Twilio, the 514 number your customers dial) doesn't change either. The model lives behind the curtain. It's your provider — whether that's ElevenLabs, Retell, or OpenAI's native platform we covered in our piece on OpenAI's native SIP — that handles transport. Voice is voice; phone is phone.
Verdict: migrate, wait, or watch?
It depends on where you are. Here's our honest read:
- If you're still on gpt-4o-realtime-preview or an equivalent model from 6 months ago — migrate. The performance jump plus the 20% price drop makes waiting hard to justify. Plan 30 to 60 minutes for a parallel test.
- If you're on GPT-Realtime-1.5 from a few months ago — run it in parallel for 7 days, compare the metrics (call completion rate, average duration, human-handoff rate), then decide.
- If you're on ElevenLabs Agents or Vapi — no panic, those platforms will likely add GPT-Realtime-2 as an engine option in the coming weeks. Neowin has a detailed ecosystem rundown.
For a Quebec SMB already running an AI Voice Agent with us, we're triggering a 48-hour A/B test this week. Last week's metrics as baseline, 20% of traffic on GPT-Realtime-2, and if the KPIs hold, we ramp to 100% by Friday. No crystal ball, just data.
Bottom line
OpenAI just raised the floor for AI Voice Agent quality. What was "good enough" yesterday is, as of this morning, behind. For a Quebec SMB, that means three priorities this week: (1) check which model version you're running on; (2) talk to your provider about migration; (3) don't confuse "the model is better" with "everything else follows" — compliance, integration, and prompt quality remain human work.
And if you don't have an AI Voice Agent yet, this is probably the moment when the gap between having one and not having one starts showing up in your end-of-month numbers.
Primary source: 9to5Mac, May 7, 2026.