On May 12, 2026, Vapi closed a $50M Series B at a $500M valuation after beating 40 competitors to win Amazon Ring. Three days later, ElevenLabs added Gemini 3.1 Pro and Qwen 35-397B to its Conversational AI platform. And on May 8, OpenAI released GPT-Realtime-2 with a 128k token context window — quadrupling the memory of voice agents.
For a Quebec SMB shopping for an AI voice agent in May 2026, these three announcements redraw the landscape. The question "which platform should I pick?" doesn't look like it did in January. Here's the no-filter comparison of the three platforms any serious vendor will quote you today: Vapi, ElevenLabs Conversational AI 2.0, and Retell AI.
The three platforms in 30 seconds
Vapi. Founded 2023, San Francisco. An "orchestrator" platform: you pick your LLM (OpenAI, Anthropic, Google), your STT (Deepgram, Whisper), your TTS (ElevenLabs, Cartesia, PlayHT). Vapi handles the plumbing and bills per minute. Riding high after the Amazon Ring deal, $500M valuation.
ElevenLabs Conversational AI 2.0. The reference TTS vendor is now a vertically-integrated platform: in-house voices (multilingual v2, turbo v2.5), Scribe STT, agents with visual workflows, Git-style versioning launched May 16, 2026. LLM choices: OpenAI, Anthropic, Google Gemini 3.1, Qwen 35-397B (new).
Retell AI. Younger, more focused. Specializes in high-volume outbound agents (collections, appointment setting, lead qualification). Aggressive pricing, sub-800ms latency on average. Less flexible on custom voices than the other two.
Criterion 1 — Voice quality in Canadian French
This is where most demos collapse. A voice that sounds perfect in Parisian French can scare a customer in Trois-Rivières on the first sentence. 38% of Quebec SMBs that cancelled their deployment between January and April 2026 did so for this exact reason.
ElevenLabs wins here. Multilingual v2 handles Quebec French with a credible neutral accent, and you can clone from 60 seconds of in-house audio — perfect if the SMB wants the agent to sound like the owner or her team. Recording quality scored "studio/good/ok/poor/bad" since May 2026 (refuse anything below "good").
Vapi depends on the TTS you pick. With ElevenLabs in the backend, quality is identical. With Cartesia or PlayHT, you get weird pronunciations ("dépanneur" read as "dépaneur", "Côte-des-Neiges" massacred). Test systematically with a script of 20 Quebec street names and toponyms.
Retell AI uses mainly ElevenLabs and Cartesia. Good in English, average in Quebec French. Latency is better (650ms vs 850ms average) but the default accent stays neutral-European.
Verdict criterion 1: ElevenLabs direct or Vapi+ElevenLabs backend. Retell for English-first.
Criterion 2 — The integrated LLM brain
Since May 8, 2026, OpenAI GPT-Realtime-2 (128k tokens, GPT-5-class reasoning) is the reference for complex conversations. But Gemini 3.1 Pro and Qwen 35-397B (Alibaba's massive model, 397 billion parameters) are now reachable too.
Vapi. Truly open marketplace. You can route by complexity: Claude Haiku 4.5 for "yes/no/transfer", GPT-Realtime-2 for complex sales. Maximum flexibility, maximum configuration complexity.
ElevenLabs Conversational AI 2.0. The four major LLMs are native (OpenAI, Anthropic, Gemini 3.1, Qwen 35-397B). Switch via dropdown. You lose the ability to mix multiple LLMs in one agent (unless you split via sub-agents in multi-agent architecture).
Retell AI. OpenAI and Anthropic, period. No Gemini, no Qwen. For 80% of SMB use cases, this is fine — for anyone wanting to test GPT-5 against Qwen 397B, it's a no-go.
If LLM choice feels overwhelming, read our detailed comparison of GPT-Realtime-2 vs Gemini 3.1 vs Claude Sonnet 4.6 vs Qwen 35 before signing anything.
Verdict criterion 2: Vapi for total optimizers. ElevenLabs for the broadest LLM palette in a single interface. Retell for those who know OpenAI is enough.
Criterion 3 — Law 25 compliance and Canadian hosting
This is the criterion that eliminates half the candidates. Law 25 (Quebec) and PIPEDA (federal) require personal information to be hosted or processed under specific rules, with documented consent and deletion mechanisms.
Vapi. Main hosting in the US (AWS us-east). No Canadian region available in May 2026. Cross-border transfers require a proper Privacy Impact Assessment (PIA). Doable, but not a "30-minute setup".
ElevenLabs. Now offers an "EU/CA residency" option (announced April 12, 2026) with hosting in Germany or Canada-Central (Toronto) for Business and Enterprise clients. Simplest path to Law 25 compliance.
Retell AI. US hosting only, no announced CA roadmap. For Quebec health clinics, law firms, or financial services, this complicates compliance analysis.
For the audit details every vendor should hand you, see our 6-point Law 25 audit. And yes, TECHMA handles this analysis for its clients — not the client trying to decode OpenAI's docs.
Verdict criterion 3: ElevenLabs wins clearly. Vapi possible with legal support. Retell: avoid for regulated sectors.
Criterion 4 — Real costs at Quebec SMB volume
A typical Quebec SMB (retail, clinic, home service) receives between 400 and 2,000 calls/month depending on size. Let's run the numbers for 800 calls/month, average 3:30 duration, so ~46 voice hours/month.
Vapi. Per-minute billing: ~$0.12 USD/min agent + LLM + TTS pass-through. At 46 hours, about $540 USD/month with GPT-Realtime-2 + ElevenLabs. High variability by LLM.
ElevenLabs Conversational AI 2.0. Pro plan all-inclusive: $0.15 USD/min in French, $0.12 USD/min in English. At 46 hours, about $360 USD/month. No billing surprises.
Retell AI. Cheapest: ~$0.07 USD/min, about $200 USD/month. But hidden costs on SMS follow-ups, long-term recording storage, premium voice cloning.
That said, monthly agent cost is nothing next to $126K/year of missed calls for an average SMB. The question isn't "which is cheapest", it's "which recovers the most lost revenue for the least deployment friction".
Verdict criterion 4: Retell for tight budgets without cloning. ElevenLabs for predictable billing. Vapi for ounce-by-ounce optimizers.
Criterion 5 — Tools, webhooks, and native integrations
A voice agent that doesn't plug into your booking system, your CRM, or your SMS is an agent that adds no value. The "integrations" criterion plays on three axes: webhooks, native integrations, and SMS bridge.
Vapi. Full webhooks, advanced tool calling, native integrations with Make, Zapier, n8n. SMS Twilio bridge to configure manually. Excellent for developers.
ElevenLabs Conversational AI 2.0. Native integrations with Cal.com, Twilio SMS, Make, n8n. Tool calling supported. The visual workflow makes configuration accessible without a dev — fastest deployment for an SMB that wants to "see the agent calling" in a few days.
Retell AI. Webhooks and tool calling. No native Cal.com integration, goes through Zapier. More "API only" oriented, less ergonomic for SMBs wanting a supervision dashboard.
The multi-agent architecture that cuts call abandonment from 67% to 8% is now native at ElevenLabs (sub-agents with branch_id), demanding to orchestrate at Vapi, and complex to replicate at Retell.
Verdict criterion 5: ElevenLabs for production speed. Vapi for orchestration depth. Retell for API-first.
Final verdict by Quebec SMB profile
Retail, restaurant, home service (1 to 25 employees, mainly FR-CA): ElevenLabs Conversational AI 2.0. Credible Quebec voice, available CA hosting, visual workflow, predictable pricing make this the default choice.
Health clinic, law firm, financial services (Law 25 critical): ElevenLabs with CA-Central residency + documented Law 25 audit. Vapi possible but adds 30 to 60 days of legal analysis.
Outbound call center, collections, massive qualification (10K+ calls/month): Vapi, because fine-tuning by LLM cuts 15 to 25% of monthly bill at volume. Retell as second choice if English dominates.
Bilingual SMB with 50% EN / 50% FR-CA clientele: ElevenLabs (native language switching managed without a hack) or Vapi with GPT-Realtime-Translate (automatic switching across 70 languages).
One thing the three platforms will never tell you
None of the three works "on its own". The platform is 30% of the deliverable. The remaining 70% is analyzing your call flows, writing the system prompt, configuring CRM tools, testing on 50 real Quebec scenarios, data residency, and daily conversation monitoring during the first 60 days.
At TECHMA, we handle this 70% for our clients. The client picks the platform with us based on profile (see verdict above). We take charge of Zoho/HubSpot/Cal.com integrations, Law 25 compliance, bilingual testing on Quebec scenarios, and post-launch support. That's what turns "I signed for an AI voice agent" into "my missed calls dropped 38% in 60 days".
Actively shopping? Before signing with any single vendor, demand a 15-minute bilingual demo with a script from your real use case. And demand the table above filled in for your specific SMB profile. That conversation makes the difference between a 2026 deployment that succeeds and an account abandoned in 90 days.