Tuesday, 2:12 p.m. The owner of an auto shop in Sherbrooke stares at the transcript of the last call and grumbles. The customer had asked whether he could replace his summer tires with the D24s he had heard about, since his 2018 Camry was due for its annual inspection. The AI voice agent answered in four seconds — fluent, natural, articulate. But it recommended the D24s, which only fit 19-inch rims, while the 2018 Camry leaves the factory on 17-inch wheels.
The problem was not the language model. The problem was that the agent did not take the time to reason. It answered fast, period.
That scenario — common across Quebec for the last eighteen months — is about to change. On May 7, 2026, OpenAI shipped GPT-Realtime-2, its first voice model capable of reasoning while it speaks. And along with it, a five-position dial that most Quebec SMBs don't yet know they should be using.
What actually changed on May 7, 2026
For two years, building an intelligent voice agent meant stacking three technologies: STT (speech-to-text), LLM (reasoning), TTS (text-to-speech). That chain accumulated latency at every link. A round trip to GPT-4o in the middle added 800 ms to 1.4 s. A round trip to a reasoning model like o3 added two to four seconds. Past the 700 ms threshold that separates a professional voice agent from an impostor, the customer started asking "hello?".
That's why voice agents until now have systematically been dumb but fast. Nobody wanted to lose the conversation over ten seconds of silence.
GPT-Realtime-2 collapses that chain. Reasoning now lives inside the audio loop, with no text detour. The context window jumps from 32,000 to 128,000 tokens — a full one-hour call remains in memory. And, crucially, OpenAI exposes five reasoning levels that the developer can pick per intent.
Per OpenAI's official documentation, GPT-Realtime-2 scores 15.2% better than GPT-Realtime-1.5 on the Big Bench Audio benchmark. Zillow, the first named customer, reports a 26-point lift in call success — 95% versus 69% before.
The five-position dial: what each level actually does
Here is how the five levels behave in practice, with the orders of magnitude observed over the last two weeks.
Minimal. No internal reasoning. The model answers in reflex mode. Added latency: 200 to 400 ms. Ideal for confirmations, greetings, simple transfers. Cost equivalent to GPT-Realtime-1.5.
Low (default). Short reasoning, basic checks. Added latency: 400 to 700 ms. The right default for 70% of calls at a Quebec SMB — bookings, basic questions, redirects.
Medium. The model plans in several steps before speaking. Added latency: 700 to 1,100 ms. Useful for triaging an emergency, disambiguating a request, choosing between two products.
High. Extended reasoning, cross-checking. Added latency: 1,100 to 1,800 ms. Reserve for cases where a wrong answer is expensive — financial qualification, first-line legal advice, technical diagnostic.
Xhigh. Deep deliberation, comparable to o3-class reasoning under the hood. Added latency: 2 seconds or more. Use exclusively for multi-step transactions with cross-constraints. On a consumer-facing voice agent, it's almost always a mistake.
Why the speed-vs-intelligence trade-off just died
The point is not that GPT-Realtime-2 reasons better. The point is that you can choose how hard it reasons, request by request. The Heyloha team, which has been tracking the OpenAI API since March 2026, captures the break well: "what used to require a brittle three-link pipeline now fits into a single end-to-end model".
The concrete result: on a typical Quebec auto-shop call — qualifying a tire request — the agent can open in low ("hello, Sherbrooke Mechanic, how can I help you?"), shift to medium the moment the customer mentions the car (verify rim/tire compatibility in memory), and come back to low to confirm the appointment. Three levels in a single call.
This dynamic did not exist before May 7. The developer picked a model and that was it.
Which level for which Quebec SMB
The practical question every TECHMA client is asking this week: "where do I start?". Here is what reading transcripts from 4,000 Quebec calls over the last six months suggests.
A hair salon in Lévis that mostly takes appointments: minimal + low. The phrasing is simple, the intent clear. Pushing to medium would slow things down with no gain.
A dental clinic in Trois-Rivières that hires an AI voice agent before a hygienist: low by default, medium on insurance questions. Dental emergency triage demands planning — not a reflex answer.
A Montreal restaurant dealing with no-show bookings: low almost everywhere. The conversation is scripted, exceptions are rare.
An independent auto shop in Saguenay: low for greeting, medium to qualify the vehicle and check parts compatibility. The accuracy gain easily covers the extra 400 ms.
A Quebec City law firm screening after-hours calls: medium as standard, high the moment the caller mentions an urgency or a legal deadline. The cost of mis-prioritization is too high.
An accounting firm during tax season: high on qualification questions, low on booking. The client prefers waiting 1.5 s for the right answer over hearing wrong advice in 400 ms.
The math for a typical Quebec SMB
Let's run the numbers on a real case: a dental clinic taking 500 calls per month, average length 2 minutes 50, for about 1,415 total minutes.
OpenAI pricing for GPT-Realtime-2: $32 USD per million input audio tokens, $64 USD per million output audio tokens. A typical call burns roughly 6,000 input tokens and 4,000 output tokens, so $0.192 + $0.256 = $0.448 per call at low. For 500 calls, about $224 USD per month in model cost.
Moving 30% of calls to medium (the insurance questions and emergency triage) adds roughly 25% to output token consumption, or about $17 USD more. For an extra $17, the error rate on insurance questions drops from 12% to 3% based on the transcripts we have analyzed at our dental clients.
Pushing high indiscriminately, on the other hand, would double the bill to $450 USD per month — with no measurable improvement on the 70% of simple calls. That is exactly the trap that the hidden bill of $1,847 per month documents: over-configuring blows up operating cost.
For comparison, a human agent working peak hours at $22/hour for 30 hours per month costs $660 — with no after-hours coverage.
The contradiction nobody articulates
More reasoning is not better. It is often worse.
A xhigh setting on a 5 p.m. haircut booking adds two seconds of silence, makes the customer doubt ("are you still there?"), and costs three times more than a low. On the transcripts where we forced high on simple calls, abandonment rose from 4% to 11%.
The Quebec customer is not expecting a philosopher on the other end. They expect a competent assistant who answers fast when it is simple and who visibly thinks when it is complicated. The key word is visibly.
That's where preambles come in — another May 7 novelty. The model can now say "one moment" or "let me check that" while it reasons in high. Combined with parallel tool calls that include audio feedback ("I'm pulling up your patient record"), the 1.5-second latency at high becomes invisible: the customer perceives work, not silence.
Why this dial changes the Law 25 and quality calculus
A detail often overlooked: the higher the reasoning level, the more the agent correctly refuses out-of-scope requests. On the internal tests we ran since May 12, the correct refusal rate on mental health questions ("I'm not feeling well, what should I do?") rises from 78% at low to 96% at medium. That is a serious improvement on the Law 25 compliance side and the 9-step protocol.
Conversely, xhigh has a side effect: it sometimes tries too hard to answer questions it should refuse. "Smarter" turns into "too confident". It is counter-intuitive, but it shows up in the transcripts.
Per the April 2026 latency benchmarks compiled by Trillet, moving from low to medium costs an average of 400 ms — exactly the budget that remains before crossing the 700 ms threshold. That is why per-intent configuration, not global default, has become the only viable approach.
What TECHMA actually configures for Quebec SMBs
Here is the pattern we have applied since May 10 on every agent we deploy.
First link: a lightweight, near-free intent classifier that tags the customer's opening line. "Hello, I'd like to book an appointment" → minimal. "I have a complicated insurance question" → medium. "My child broke a tooth, what do I do?" → high.
Second link: a conservative default level (low in 90% of cases), with dynamic escalation on explicit flags — mention of urgency, internal contradiction, transfer request.
Third link: a human fallback above high. If the customer is pushing toward xhigh, it is probably too complex for voice anyway — we transfer.
This is not a setup the SMB does itself. The entire ecosystem — from intent design to continuous transcript analysis — is part of the deployment and monitoring package TECHMA delivers on the client's behalf.
What this means for you, Quebec SMB owner
You don't have to understand the dial. You have to understand that the quality of your voice agent is no longer limited by the model available. It is limited by the quality of the configuration that someone does for you.
If you deployed a voice agent before May 7, 2026, it is probably still running on GPT-Realtime-1.5, with a fixed reasoning level. Your transcripts probably show the same kinds of errors as the Sherbrooke auto shop at the top of this article: quick answer, missing reasoning.
If you are considering deploying one, the right reflex is to ask your integrator: which reasoning level, by intent, and at what projected monthly cost? The answer will immediately tell you who you are talking to.
Book a free audit of your voice agent — we analyze your last 100 calls, identify the under-reasoned intents, and price the cost and quality gap that migrating to GPT-Realtime-2 would give you, level by level.
The dial has been live for two weeks. The first gains go to the SMBs that configure it first.