Two weeks ago, a TECHMA client — a South Shore dental practice with three simultaneous phone lines — called his own number from a Tim Hortons in Brossard. Not to test. To validate. Coffee in hand, he was timing the delay between his question and the first syllable of his AI voice agent's reply. 620 milliseconds. He smiled, hung up, and texted the number over.
That 620ms, in May 2026, has become a political statistic. Because between May 7 and May 12, two announcements shifted the ground under the entire voice agent industry: OpenAI launched GPT-Realtime-2 with GPT-5-class reasoning and a 128,000-token context window, and ElevenLabs pushed SDK v2.47.0 with new LLM options (gemini-3.1-pro-preview, qwen35-397b-a17b) directly pluggable into agents. The bar for what counts as "conversational" just moved.
The new threshold nobody talks loudly enough about is 700ms. Above it, your SMB loses calls. Below it, you're in the race. And most of the vendors pitching Quebec SMBs right now are above — often well above. This article gives you the raw May 2026 numbers, compares them across platforms, and explains why a bilingual FR-EN voice agent in the Quebec context is even more latency-sensitive than people think.
What actually happened between May 7 and May 12, 2026
Two announcements, five days apart, the same effect on the market.
May 7, 2026 — OpenAI ships three new voice models. GPT-Realtime-2 (GPT-5 reasoning, 128k context, $32/M input tokens, $64/M output), GPT-Realtime-Translate (real-time translation from 70 languages to 13), and GPT-Realtime-Whisper (streaming transcription at $0.017/minute). The detail that flipped many architects: GPT-Realtime-2 can call multiple tools in parallel and announce those actions out loud ("checking your file…"), which masks perceived latency surgically. The official announcement is here.
May 12, 2026 — ElevenLabs pushes v2.47.0. Phone number agent configuration directly via API, custom webhook headers, RAG chunk listing, and most importantly — gemini-3.1-pro-preview and both Qwen 35 models (35B and 397B) added to the list of LLMs available in agent config. For a Quebec deployer, that means you can now test a non-American "brain" on an agent that speaks Québécois without changing platforms. The full changelog is here.
The two announcements share one thing: they attack the same bottleneck — latency. Not voice quality. Not price. Latency.
Why 700ms, not 500ms or one full second?
You have to go back to psycholinguistics research to understand. In natural human conversation — regardless of language, regardless of culture — the median gap between the end of one phrase and the start of a reply is about 200 milliseconds. It's wired into the brain. When someone consistently exceeds 400ms, we start to perceive them as distracted, hesitant, or just "slow". Above 800ms, contact centers report 40% more hang-ups than the same conversation under the threshold.
700ms is the pragmatic ceiling: it's the upper bound that keeps a voice agent in the zone where the caller doesn't actively notice they're talking to a slow machine. Hamming AI documents this threshold precisely in their 2026 technical guide. Above it, you accumulate micro-frustrations that add up — exactly the kind of friction we already dissected in our analysis of the 7 pre-hangup frustration signals.
The raw scoreboard — who delivers what in May 2026
Here are median figures from 1,200+ independent test calls published this spring:
- ElevenLabs Conversational AI 2.0: 400–600ms (voice generation alone); 500–700ms full-loop with native turn-taking
- Vapi (optimized stack — Deepgram + GPT-4o-mini + ElevenLabs Turbo + Twilio): 500–700ms; but standard deviation is high because Vapi orchestrates 4-5 external APIs
- Retell AI: 580–620ms published, 620–800ms measured under load; proprietary turn-taking, so less jitter
- Bland AI: ~800ms — above the conversational threshold
- OpenAI GPT-Realtime-2 (May 2026): 300–500ms end-to-end; the biggest jump of the spring
Three things jump out. First, ElevenLabs and OpenAI are now the only two providers where the median is structurally below 700ms — not because of a well-tuned stack, but thanks to native architecture. Second, Vapi can reach the same zone, but only with surgical provider assembly; a poorly configured deployment drifts fast to 900-1100ms. Third, Bland AI is now out of zone — and that changes the conversation for Quebec SMBs who signed with them in 2025.
The bilingual FR-EN trap American benchmarks ignore
All the numbers above are measured on unilingual English calls. In Quebec, we have an additional problem that never shows up in Silicon Valley benchmarks: mid-sentence code-switching. "Hi, je cherche un rendez-vous, can you check Wednesday afternoon?" That's a typical call for a Westmount dental clinic or a Pointe-Claire law firm.
On architectures where the STT → LLM → TTS pipeline has to re-detect language on every turn, we observe in practice an additional cost of 80 to 150ms per linguistic switch. Which means an agent at 620ms in pure English can climb to 770ms on a bilingual call — above the threshold. The only providers who don't pay this tax in 2026 are the ones whose language detection model runs in parallel rather than in series. ElevenLabs Conversational AI 2.0 and GPT-Realtime-2 are currently the two that do it cleanly. We already quantified what that's worth concretely for a veterinary clinic — it's $96,000 per year recovered on emergency calls.
The dollar cost of 950ms latency
Let's stop talking in milliseconds for a moment. Concrete case. A Quebec SMB receiving 1,200 calls per month (typical volume for a mid-size medical practice, a restaurant with reservations, or a home-services company). With a voice agent at 950ms — above threshold — we expect roughly 12% more hang-ups compared to an agent at 600ms. On 1,200 calls, that's 144 lost calls per month.
If the average converted call value is around $85 (medical appointment or service booking), and the conversion rate of a normal call is 30%, then monthly loss lands at $3,672. Over the year, that's $44,064 going up in smoke because the agent has a 350ms ignition delay. That's half the salary of a full-time employee, lost because nobody timed the initial deployment.
We've run this calculation with several TECHMA clients as part of our total cost of ownership (TCO) analysis. The pattern is consistent: latency is the blind spot that turns a "$499/month, profitable" project into "$1,847/month that doesn't deliver".
What a Quebec SMB must demand from its vendor in 2026
If you're evaluating an AI voice agent this month — or already have one in production — here's the new minimum list:
- Median under 700ms under load, measured on 50+ real calls, not in a controlled demo
- Bilingual FR-EN median under 800ms, measured on real code-switching (not just a French call followed by an English call)
- Native architecture (turn-taking integrated into the model, not chained via external VAD)
- Contractual commitment on latency P95, not just the median — it's the P95 that bites you in peak hours
- Ability to test GPT-Realtime-2 or ElevenLabs Conv. AI 2.0 without a full migration — a vendor that locks you to a single model in May 2026 leaves you behind by July
This list isn't theoretical. It comes from five TECHMA deployments in the last three months where we replaced a vendor delivering 1,100ms with an ElevenLabs 2.0 + Law 25-compliant stack delivering 580ms. If you want to know how to test a vendor before signing using the 7-call protocol, we've documented that in detail too.
What TECHMA does with this data for its clients
All integrations between your AI Voice Agent, your CRM, your PBX, and your appointment system are deployed by our team — not self-serve. When we measure 580ms at a client site, that number is defended by an architecture we chose line by line: ElevenLabs Conv. AI 2.0 for the conversational engine when bilingualism is primary, GPT-Realtime-2 when reasoning is primary, and a custom router to switch between them based on the intent detected in the first second.
If you're currently running on misconfigured Vapi, on Bland, or on a homegrown solution drifting above the threshold, we can run a latency audit in 48 hours, measure where you are, and estimate real annual loss. Contact us or consult our full Vapi / ElevenLabs / Retell comparison for Quebec SMBs before your next decision.
700ms isn't marketing fluff. It's a biological floor that found itself a calendar in May 2026. Picking a vendor that doesn't respect it in mid-2026 means choosing to lose 12% of your calls in silence.
One more wrinkle: the P95 problem nobody benchmarks
Median latency is the comfortable number every vendor will quote you. But it's the wrong number to negotiate on. What kills SMB deployments is the P95 — the latency that 5% of your calls experience. If your median is 620ms but your P95 is 1,400ms, then 60 calls per 1,200 will land in pure "machine sounds slow" territory. And those 60 calls — guess what — are disproportionately your most valuable ones, because peak hours are when load spikes.
The new May 2026 architectures help with P95 specifically. GPT-Realtime-2's parallel tool calling means that a CRM lookup that used to add 400ms in series now adds zero — because the model speaks "checking your file" while the lookup runs. ElevenLabs Conv. AI 2.0's native turn-taking means VAD jitter no longer compounds across the conversation. Both architectures attack P95, not just the median. That's the part most resellers don't explain because they don't measure it themselves.
If your current vendor can't quote you a P95 number, that's not because the data doesn't exist. It's because they haven't measured it — which usually means it's bad. Demanding P95 in your contract is, in May 2026, the cheapest insurance policy a Quebec SMB can buy.