OpenAI Realtime vs ElevenLabs Agents: Choosing a Voice Architecture

OpenAI Realtime and ElevenLabs Agents are not simply two voices to swap. The first comparison is between a realtime model/API and an agent platform. A second OpenAI option, GPT-Live, separates the voice interaction from a delegated backend. Those boundaries affect what you build, what you configure and what you must test.

This guide was reviewed against official documentation on September 25, 2026. It is an architecture comparison, not evidence that one service is faster or cheaper on the same production calls.

Understand the three options

Read the Realtime overview, GPT-Live migration guide and ElevenAgents overview. The current documentation matters: do not apply a tutorial for one architecture unchanged to another.

GPT-Live changes where the instructions belong

OpenAI distinguishes Responses delegation, where a hosted backend reasons about tasks, from client delegation, where your application retains its orchestrator. In both arrangements, the application still owns custom function execution and permission checks. Its migration guidance separates voice behavior from detailed business procedures. OpenAI's delegation guidance

For a receptionist, the voice instructions might cover concise speech and how to handle caller interruptions. The backend instructions cover which business facts are approved, what an appointment request requires and when a booking may be created.

Changing the speech architecture does not make a calendar connection valid, stop duplicate writes by itself or establish permission to call someone. Keep those responsibilities explicit when estimating implementation work.

Compare business outcomes before voice style

Use one scenario in each system: a caller asks a question, provides contact details, changes one detail and requests a next step. Keep the supplied business facts the same.

Record four separate results: what the caller said, what the agent understood, what action the application attempted and what the external system actually recorded. For a booking, a spoken “confirmed” without a returned successful booking is a failed test.

Add an unavailable slot, an external timeout and a correction while the agent is speaking. Verify that a late result for the old request does not become the answer to the new one. These tests expose problems that a polished greeting will not reveal.

Compare ownership and cost

Ask which parts the platform handles and which your team must operate: telephone setup, number lifecycle, authentication, tool execution, transcripts, recording permissions, retention, support and usage reconciliation. Even when a product supplies a convenient default, verify that it fits your required deployment.

For costs, use one complete call configuration and its actual usage. Include audio, backend reasoning, telephone service, platform charges and recurring resources where applicable. Do not compare a model-only rate with an agent-platform total.

Where Burki fits

Burki uses LiveKit for its voice runtime and offers GPT Live as a voice mode. The value to assess is the surrounding create, test, review and publish workflow, plus the specific connected actions your workspace can use. This does not imply Burki supports every OpenAI or ElevenLabs feature or every delegation combination.

Start with an unpublished receptionist draft, review the instructions and inspect the result of a browser practice conversation. Then verify the actual telephone and action paths you plan to deploy. The demo checklist keeps those acceptance steps separate.