On April 7, 2026, ElevenLabs slipped a quiet line into its changelog: "Gemini 3.1 Pro Preview, qwen35-35b-a3b and qwen35-397b-a17b added to the list of LLM providers." Nobody really talked about it in Quebec. Yet for a small business currently paying ~$8.45/hour of call on GPT-Realtime, this update potentially divides the bill by 7 — or even by 12. As long as you pick the right option, and you understand what really changes behind the voice.
As integrators on the ground in Quebec, we ran all three LLMs in production on a real voice agent for three weeks. Here's what we saw — numbers, latency, Quebec French quality, and the Law 25 trap nobody mentions.
TL;DR — The 4-line verdict
GPT-Realtime (OpenAI) → rock-solid stability, but ~$8.45/call hour. Reserve for critical use cases.
Gemini 3.1 Pro (Google) → 7 to 12× cheaper, montreal-1 hosting available. Best fit for 90% of SMBs.
Qwen 35B (Alibaba) → unbeatable on cost (~$0.70/hour), surprisingly strong on raw quality. Red flag on Law 25.
Our default pick for Quebec clients in May 2026: Gemini 3.1 Pro.
Quick comparison
Now, let's dive into each option. Not in marketing mode — in integrator mode, having seen each model crack or hold up under real load.
GPT-Realtime: the Mercedes of the phone line
OpenAI moved its Realtime API to GA in late 2025, and since then it's been the quiet default for production voice agents. The gpt-realtime model is their best speech-to-speech to date — audio in, audio out, no intermediate text step that wastes time.
From the field:
- Stable latency around 350 ms in steady state. The model "holds the rhythm" even on long 8-10 minute conversations.
- Impeccable turn-taking. OpenAI's built-in VAD (voice activity detection) handles interruptions, hesitations, and "uhh" sounds well. This is still where OpenAI dominates.
- Quebec French: excellent. GPT-Realtime handles "asteure", says "514 555-2113" naturally, and pronounces Quebec last names (Tremblay, Bélanger, Dufour) with rare accuracy. Not perfect, but clearly the best of the three.
The problem is the price. OpenAI charges roughly $40 USD per million audio input tokens and $80 USD output. A voice conversation burns about 2,200 tokens per minute. Brutal math: $8.45 USD per call hour. For a clinic taking 200 hours of calls per month, that's $1,690 USD just for the LLM — before telephony, before ElevenLabs TTS, before your integrator.
The other big sticking point is hosting. OpenAI routes to its US datacenters. For a Quebec law firm dealing with information protected by attorney-client privilege — or for a legal practice trying to honor its Law 25 obligations — that complicates things. You can write a PIA (privacy impact assessment) that passes muster, but it's extra paperwork.
Pick it if: your use case is critical (medical urgency, 24/7 technical support, banking), your budget tolerates $1,500-2,500/month for the LLM alone, and you want the most mature SDK with the best documentation. See OpenAI's official gpt-realtime announcement.
Gemini 3.1 Pro: the best price/quality ratio of 2026
Google released Gemini 3.1 Pro in preview in early 2026, and ElevenLabs added it as an LLM option for its agents on April 7. In our view, it's the most important change of the year for Quebec SMBs.
Three reasons:
- Cost. Gemini 3.1 Flash Live charges roughly $3/M input tokens and $12/M output. On the same call hour, that comes out to ~$1.11 USD. For the 200 h/month clinic mentioned earlier: $222 USD/month instead of $1,690. That's about $1,470 in monthly savings, just on the LLM.
- Generation speed. Gemini Flash outputs around 250 tokens per second, vs ~77 t/s for Claude Sonnet. In a voice conversation where the LLM accounts for 71% of the latency budget (per ElevenLabs benchmarks), that speed concretely changes how the call feels. The customer doesn't get the impression they're talking to a hesitant machine.
- Data residency. And this is huge. Google Cloud offers a montreal-1 region for Gemini Pro. Concretely: your voice agent's requests stay in Canada. For Law 25 compliance, that's the killer argument. No other LLM offers it today.
The catch — there's always one — is Quebec French. Gemini 3.1 Pro was trained on universal French. It says "voiture" instead of "char", "60 dollars" instead of "60 piastres". On Quebec proper nouns, we measured about 8% pronunciation errors on a sample of 200 last names (vs 3% for GPT-Realtime). That's manageable with a system prompt that forces the register, but it's not plug-and-play like OpenAI.
Another detail: TTFT (time to first token, i.e. the delay before the model starts replying) is 960 ms. So the very first sentence of the conversation has a noticeable half-delay. Once the rhythm sets in, you fall back to 350 ms — but you need to know it, and adjust your greeting accordingly.
Pick it if: you want the best possible ROI, you have more than 50 hours of calls per month, and Law 25 compliance is a real concern. It's our default pick for new Quebec clients since April 15, 2026.
Qwen 35B: the dark horse (with an asterisk)
Qwen is Alibaba's family of LLMs. Until recently, most Western integrators ignored them — not because they were bad, but because nobody talked about them. The April 7 ElevenLabs changelog changed that.
The two added models are:
- qwen35-35b-a3b: 35 billion parameters, 3B active (mixture of experts). Pricing around $0.50/M input, $1.50/M output. Call hour: ~$0.70 USD.
- qwen35-397b-a17b: the big version. 397B parameters, 17B active. Higher pricing (~$3/M and $9/M), but better quality.
Benchmark surprise: on the seed-hard test (transcription under hard conditions), Qwen 3.5-Omni Plus posted a WER (word error rate) of 6.24 vs 27.70 for ElevenLabs in native mode. Four times better. That's the kind of number that raises an eyebrow the first time you see it.
What we saw ourselves on Qwen 35B in production:
- Latency: 400-700 ms. More volatile than the other two. On Alibaba's European servers, we observed spikes up to 1,200 ms at Asian peak hours.
- Standard French: surprisingly correct. The model has clearly seen a lot of Francophone data.
- Quebec French: validate seriously. On our sample, ~15% pronunciation errors on typical Quebec last names. "Gauthier" sometimes came out as "Go-thi-ay". Not catastrophic, but worth testing.
- Tool calling: still a bit erratic. On 100 calls where the LLM had to trigger a function (booking, customer lookup), we saw 4-5% silent failures.
But the real problem is elsewhere. Read the next section.
The Law 25 trap nobody will tell you about
The choice of LLM directly impacts where your customers' data lives. And Law 25 (Quebec's modernized personal information law), since its full enforcement in September 2024, requires businesses to document, justify, and minimize transfers of personal information outside Quebec.
Concretely:
- GPT-Realtime routes data to the United States. Subject to the US CLOUD Act. For most Quebec SMBs, this stays manageable with a tightly-written PIA. For healthcare practices or law firms, it's a sensitive point that needs to be documented.
- Gemini 3.1 Pro via Google Cloud Montreal-1: data stays in Canada. Not in Quebec strictly speaking, but on Canadian soil. This is currently the best compromise on the market. The CAI considers Canada a "comparable" jurisdiction for transfers.
- Qwen 35B via Aliyun: main datacenters in Singapore, Hong Kong, and mainland China. For a Quebec SMB, that means transferring personal information to a jurisdiction that is not recognized as "comparable" by the CAI. A Quebec law firm cannot, in practice, use Qwen for client call traffic without substantially overhauling its data governance regime.
Bottom line: Qwen is technically attractive. Legally in Quebec in 2026, it's complicated for regulated industries. For a retail shop or restaurant, much less of a concern.
4 concrete Quebec SMB scenarios
1. Dental clinic in Trois-Rivières — 90 hours of calls/month
Typical case of a 4-dentist clinic. Patient data = sensitive health information. Recommendation: Gemini 3.1 Pro on Montreal-1. Saves ~$675/month vs GPT, and clean Law 25 compliance.
2. Restaurant in Montreal — 40 hours of calls/month
Reservations, menu questions, hours. No sensitive data. Recommendation: Gemini 3.1 Pro by default, or Qwen 35B if you want to test for an extra ~$16/month savings (low stakes).
3. Garage in Sherbrooke — 60 hours of calls/month
Booking, ballpark estimates, repair follow-up. Recommendation: Gemini 3.1 Pro. Standard Quebec French is sufficient, and the ROI vs GPT-Realtime is obvious.
4. Law firm in Quebec City — 70 hours of calls/month
Attorney-client privilege, confidential client files. Recommendation: GPT-Realtime documented with PIA, OR Gemini 3.1 Pro via Montreal-1. Not Qwen, until data residency is clarified. See our 7 pre-launch tests guide.
Our recommendation by SMB size
- Small (<50 h calls/month): Gemini 3.1 Pro. The cost gap with Qwen is marginal ($15-20/month) and quality justifies the stability.
- Medium (50-300 h/month): Gemini 3.1 Pro with Montreal-1 hosting. The absolute sweet spot in 2026.
- Large (>300 h/month or multi-site): hybrid architecture. Gemini 3.1 Pro for 80% of traffic (volume), GPT-Realtime for the 20% of critical cases (escalations, urgencies, VIPs).
If you're calculating your real voice AI agent ROI, the LLM choice can represent 40-60% of your total monthly cost. It's not a technical detail — it's a financial decision.
FAQ
Can you switch LLMs mid-flight?
Yes. On ElevenLabs Agents, the LLM is changed in the agent's config in a few clicks. No prompt migration, no history loss. We do it regularly with our clients when a new model ships. The main work is re-testing the critical flows (booking, transfer, etc.) because every LLM has its little behavior quirks.
Which one speaks Quebec French best?
GPT-Realtime, hands down. Gemini 3.1 Pro is solid but stumbles on Quebec proper nouns. Qwen 35B is decent on standard French but needs serious testing before production. If you handle a lot of names (clinics, practices, home services), a system prompt that includes a list of the 200 most common Quebec last names helps significantly — regardless of LLM.
What does a complete voice agent really cost per month in 2026?
For 100 h of calls with Gemini 3.1 Pro: ~$111 USD LLM + ~$80 USD ElevenLabs TTS + ~$20 USD Twilio telephony + platform fees. Total: ~$210-280 USD/month for the technology. Our integration and maintenance package adds on top depending on scope.
And Law 25 in all this — is it just paperwork?
No. The CAI handed out its first SMB sanctions in 2025 for failure to PIA cross-border transfers. It's no longer theoretical. That's exactly why we document the LLM choice in every deployment, and why we favor Gemini Montreal-1 when the client handles sensitive information. It's not over-zeal — it's reasonable caution.
In summary
On April 7, 2026, ElevenLabs democratized the LLM choice in voice agents. For a Quebec SMB, that means three things: 1) GPT-Realtime is no longer the default by default; 2) Gemini 3.1 Pro with Montreal-1 is the best quality/cost/compliance compromise of 2026; 3) Qwen is interesting to watch, but not yet ready for regulated industries in Quebec.
At Agent IA Vocal, we deploy, configure, test, and maintain the voice agent end-to-end — LLM choice included, based on your industry and Law 25 constraints. No DIY. If you want a costed recommendation for your call volume and your sector, that's what we do every day.