Three voice platforms dominate AU AI voice builds in 2026. They're close enough in quality that the choice is rarely about voice realism — it's about latency, integration friction, and cost at scale. Here's how we actually pick.
The TL;DR
- ElevenLabs — best voice quality and widest voice library. Our default when voice naturalism matters most.
- Cartesia Sonic — lowest latency (~90ms first-token). Our pick when conversational latency is non-negotiable.
- OpenAI Realtime API — end-to-end voice + reasoning in one API. Lowest-friction setup; useful when you want to ship fast.
The differences that actually matter
Voice quality
For naturalism, ElevenLabs is still the leader in 2026 — particularly for AU-English. Cartesia is close but tends to feel slightly more "model-like". OpenAI Realtime is improving fast but is roughly 6–12 months behind ElevenLabs on subjective realism.
Latency
This matters more than most people realise. Above 600ms of response delay, the conversation feels broken; above 1.2 seconds, callers think you've hung up.
| Platform | Typical first-token latency |
|---|---|
| Cartesia Sonic | ~90ms |
| OpenAI Realtime | ~300–500ms |
| ElevenLabs Streaming | ~400–700ms |
For most business calls, ElevenLabs is fast enough. For high-stakes conversational use (sales objection handling, emergency triage), Cartesia's latency edge is meaningful.
Cost at scale
Per-minute costs in 2026, USD:
| Platform | Cost per minute (USD) |
|---|---|
| OpenAI Realtime | ~$0.06–$0.10 |
| ElevenLabs | ~$0.04–$0.08 |
| Cartesia | ~$0.03–$0.05 |
For a business handling 1,000 calls/month at 90 seconds average, that's $45–$150/month. Not the bottleneck.
Architecture complexity
OpenAI Realtime bundles STT + LLM + TTS in a single API. Simplest architecture. Trade-off: less control over individual layers.
ElevenLabs / Cartesia are TTS-only. You bring your own STT (usually Deepgram) and LLM (Claude or GPT). More control, more configuration.
Decision matrix
| Your priority | Recommended |
|---|---|
| Voice realism above all (luxury hospitality, premium B2B) | ElevenLabs |
| Latency above all (sales, emergency triage) | Cartesia + Deepgram + GPT/Claude |
| Time-to-ship (MVP, internal pilot) | OpenAI Realtime |
| Multilingual (multiple AU + Asian languages) | ElevenLabs |
| Cost at high volume (>5,000 calls/month) | Cartesia |
| Voice cloning from your staff | ElevenLabs |
What we actually use
For most client builds: ElevenLabs for voice, Deepgram for STT, Claude or GPT-4 for reasoning, Twilio for telephony. When latency is critical, swap ElevenLabs for Cartesia. For tight-deadline MVPs, OpenAI Realtime end-to-end.
Why architecture matters more than vendor
The voice vendors are converging in quality. Build vendor-neutral — every component swappable in a config change — and the choice you make today becomes a choice you can revise in six months without rewriting your application.
Key takeaways
- ElevenLabs: best realism, our default.
- Cartesia: lowest latency, pick for time-critical conversations.
- OpenAI Realtime: simplest architecture, fastest to ship.
- API costs are negligible at SMB volume — focus on integration quality.
- Build vendor-neutral so today's choice isn't tomorrow's regret.
Want a voice agent like this for your business?
30-minute discovery. We'll spec your build and send a demo within one business day.