On October 20, 2026, ElevenLabs shipped two new voice models and almost everyone ran the same headline: half-second latency. The figure quoted everywhere was 0.57 seconds — the time it takes the faster model, Eleven Realtime v1 Max, to start talking. It is a good number. It is also, I would argue, the least interesting thing about the launch.
The detail that actually matters to an owner running a clinic in Toronto, a trades shop in Calgary or a law office in Halifax sits a few paragraphs down: the new Eleven Realtime API speaks OpenAI's protocol. Not loosely. Literally — the same events, the same WebSocket plumbing that developers already wrote for OpenAI's voice stack. One engineering decision, and the question every Canadian business should ask before committing to an AI voice agent quietly changed.
The number everyone quoted was the wrong one
Speed in this category stopped being a real differentiator a while ago. We have written before about the thresholds that matter — the response-time line past which callers simply hang up — and the honest truth is that most serious platforms crossed it months ago. Eleven Realtime clocks in at 0.57 seconds for Max and 0.62 for the lighter Mini. Fast. But a rival half a step behind is not why a Montreal dentist loses a patient, and a rival half a step ahead is not why they keep one.
So when a launch leads with a latency stat, I read it as marketing that knows the audience is tired. The thing worth your attention is almost never the headline number. This time it was a sentence about compatibility.
What “speaks OpenAI's protocol” actually means, in plain terms
Skip the engineering for a second. Here is the version that matters to you. For the last couple of years, building a voice agent on one provider meant writing software that only knew how to talk to that provider. Switching later was not a toggle — it was a rebuild. That is the real reason so many businesses stalled: not the monthly price, but the sense that whatever you picked, you were stuck with.
Eleven Realtime ships as two models with a clear division of labour:
- Eleven Realtime v1 Mini — built for natural, lower-cost conversation. The default for straightforward calls: hours, directions, simple questions.
- Eleven Realtime v1 Max — built to reason before it answers and to handle heavy tool use: looking up a booking, qualifying a lead, walking through a multi-step request.
Both bill at roughly $0.08 per minute of connected audio, plus token usage. But the headline, again, is not the price. It is that because the API now follows OpenAI's Realtime specification, code written for OpenAI's voice SDK can be pointed at ElevenLabs by changing an endpoint and a model name. Independent write-ups put it bluntly: moving an agent across now comes down to a new address and a new label, not a new codebase.
There is a practical sweetener for a country like ours, too. The same platform reaches more than 90 languages and a library of over 17,000 voices. For a business fielding calls from customers who slide between English, French, Punjabi, Mandarin or Arabic over a single afternoon, that range matters — and because the underlying plumbing is now shared, you are not betting your entire multilingual setup on one company's roadmap holding up.
Why a swappable voice agent changes the math for Canadian businesses
Think about what that does to the decision. The fear that kept a Winnipeg accountant or an Ottawa property manager on the sidelines was never really “what if it doesn't work.” It was “what if it works, I build everything around it, and then the vendor raises prices, gets acquired, or falls behind — and I can't leave.” That fear was rational. We spent a lot of 2026 telling readers to be cautious about which vendor they marry in a fast-moving market.
Picture a physiotherapy clinic in Vancouver that went live last spring on whichever provider its integrator happened to prefer. Six months in, a competitor ships a model that handles its booking software better and costs less. Under the old rules, moving meant paying someone to rebuild the whole thing, so the clinic stays put and quietly overpays. Under a shared protocol, that same move starts to look like switching banks rather than rebuilding your house. And the mere fact that leaving is realistic is what keeps your current provider honest on price and quality.
A shared protocol lowers those walls. It does not knock them down — more on that in a minute — but it moves the industry toward what businesses already expect from the rest of their phone setup. You already assume you can keep your business number when you change carriers. The direction of travel now is that you should be able to keep your voice-agent logic when you change providers, too. For a buyer, that is a healthier market, full stop.
Changing an AI voice agent provider by swapping an endpoint and a model name
The honest part: this is not ElevenLabs “winning”
Here is where I have to be fair, because the opposite story — “ElevenLabs just took the lead” — is the one that writes itself and happens to be wrong.
On ElevenLabs' own reported benchmarks, Max is genuinely strong: it posted 98.8% on a 1,000-question spoken-reasoning test and 87.1% on tool-use tasks, with Mini well behind at 75.9% and 60.3%. But when you look at independent evaluation of the whole field, Eleven Realtime does not top it. As of launch day, competing systems like Grok Voice 3.0 and Gemini 3.8 Live finished more tasks and were preferred more often in blind listening tests. ElevenLabs landed mid-pack, not first.
And the switch is not free. The compatible API connects over WebSocket only, so anyone running the older real-time transport has real migration work regardless of the shared protocol. Drop down to the bare API and you also leave behind the conveniences that made the managed product easy in the first place — the phone hookup, the knowledge base, the guardrails, the monitoring. Independent analysts even note that the new models use a cascaded design under the hood rather than the single fused system the “speech to speech” branding implies. None of that is a scandal. It is just the fine print the 0.57-second headline skips.
One more detail worth a Canadian footnote: the two models do not offer identical data-residency options, and the heavier Max runs in fewer regions than Mini. If where your call data physically lives matters to you — and in health, legal or financial work it should — settle that question before the demo impresses you, not after.
So the correct read is not “ElevenLabs is now the best.” It is “the walls between vendors got lower, and the company that lowered them this week is not even the one leading the benchmarks.” That is better news than a horse-race winner, because it outlives any single release.
So how should a Canadian business choose now?
If swappability is the real shift, your buying checklist should change to match it. Three things I would hold firm on:
- Choose for the job in front of you, not the leaderboard. The “best” model in a benchmark table rarely matters more than whether the agent handles your actual calls, in your customers' languages, on a bad cell connection. Buy for your Tuesday afternoon, not for a press release.
- Insist on portability before you sign. Ask where your call scripts, prompts and knowledge base live, and whether you can export them. Ask to run the same calls against two providers before you commit to one. A shared protocol only helps you if your setup is not trapped in one vendor's private format.
- Avoid the long lock-in contract. In a market where the capability map is redrawn every few weeks, a two-year commitment is a bet against your own freedom to switch. Month-to-month, or a short term with a clean exit, is worth more than a discount that chains you in.
If you make one move this month, make it this: before you renew or sign anything, ask your provider the export question out loud — “if I leave in a year, what do I take with me?” The answer tells you more about your next year than any benchmark score will.
Independent voice-agent benchmark comparison across providers in 2026
What it does not change
A swappable back end does not rewrite your obligations to the people who call you. For a business serving customers across Canada's six time zones, the hard part was never the model — it was answering in the right language, at the right hour, without sounding like a hold queue. That work is still yours to get right. And for businesses operating in Quebec, none of this touches Law 96: your agent still greets and serves callers in French by default, whatever vendor sits behind it. A new protocol changes who you can switch to. It does not change who you answer to.
The bottom line
The quiet shift of 2026 is not that AI voice agents got faster. They did, and it barely matters anymore. The shift is that, for the first time, what you build on top of one is starting to travel with you. For two years the real risk of adopting a voice agent was that it felt like a one-way door. As of this month, that door opens both ways — and that is worth far more to a Canadian business than another tenth of a second.