AI agents · OpenClaw · self-hosting · automation

Quick Answer

GPT-Live vs ElevenLabs Conversational vs Retell for Voice Agents (July 2026)

Published:

The Landscape (July 12, 2026)

PlatformFull-duplexTelephony (SIP/PSTN)API today?Best forRough $/min
GPT-Live-1❌ (ChatGPT only)Consumer voice UX inside ChatGPTIncluded in plan
ChatGPT Realtime API⚠️ PartialVia 3rd partyCustom voice apps needing OpenAI models~$0.30/min
ElevenLabs Conv. AIVia 3rd partyWeb/app agents needing best voices$0.10-0.20
Retell✅ NativeOutbound/inbound phone agents$0.05-0.15
Vapi✅ NativeDeveloper-first phone agents$0.05-0.10
Bland✅ NativeHigh-volume outbound calls$0.09
Gemini Live API⚠️Via 3rd partyVoice + vision + screen shareVaries
Cartesia Sonic✅ (as TTS)Via 3rd partyUltra-low-latency TTS engine$0.02-0.05

Decision Tree

”Where does the voice live?”

  • In a phone call (inbound support, outbound sales, appointment reminders) → Retell / Vapi / Bland
  • In a web or mobile app (in-product chat, tutoring, coaching) → ElevenLabs Conversational AI or ChatGPT Realtime API
  • Inside ChatGPT (no custom app)GPT-Live (July 8, 2026 launch)
  • Multimodal (audio + vision)Gemini Live API

”How much control do I need?”

  • Full control over LLM, prompt, toolsRetell, Vapi, ElevenLabs, Realtime API (all let you plug in your own model)
  • Just want a voice with a personaBland (opinionated, less customization)
  • Just want to use ChatGPTGPT-Live

”What’s my volume?”

  • <1000 min/mo → any platform is fine, pick on feature fit
  • 10k-100k min/mo → cost matters; Bland or Vapi typically win
  • 100k+ min/mo → negotiate directly with the platform; Retell and ElevenLabs offer enterprise pricing

The Full-Duplex Baseline

As of July 2026, full-duplex is the assumed baseline for any voice-agent product. That means:

  • Interrupt handling (user cuts in, agent stops)
  • Backchanneling (agent’s “uh-huh”s during user speech)
  • Barge-in recovery (rejoining smoothly after interruption)

Any voice platform that’s still turn-based in July 2026 is behind. Every option in the table above supports full-duplex.

Latency Reality Check

Sub-500ms end-to-end (mouth-to-ear) is what feels human. Here’s what platforms actually deliver:

PlatformReal-world medianMedian floor
ElevenLabs Conv. AI350-500ms300ms
Retell400-600ms350ms
Vapi400-700ms350ms
Bland500-700ms400ms
Cartesia Sonic (TTS)100-150ms (TTS only)90ms
OpenAI Realtime API300-500ms250ms
GPT-Live inside ChatGPT300-400ms250ms

Real-world latency depends heavily on: LLM inference time (biggest variable), network hop count, whether you’re using GPT-4o-mini vs Claude Haiku 4.5 vs GPT-5.6 Luna under the hood. Cartesia Sonic is a TTS engine — pair it with a fast LLM for lowest total latency in custom stacks.

What Each Platform Is Best At

ElevenLabs Conversational AI

  • Best voices in the market as of July 2026
  • Voice cloning (30 seconds of source audio)
  • Fine-grained prosody control (pitch, pace, emotion)
  • SDK for JS, Python, Swift, Kotlin
  • Weakness: telephony requires third-party (Twilio/Vonage)

Retell

  • Native SIP / PSTN for phone-based agents
  • Warm transfer to humans, IVR trees, TCPA compliance
  • Salesforce, HubSpot, Zapier connectors
  • Call analytics (sentiment, keywords, script adherence)
  • Weakness: less voice customization; usually pairs with ElevenLabs

Vapi

  • Developer-first — clean SDK, MCP tool support, web hooks
  • Bring your own LLM, TTS, STT
  • Reasonable telephony via Twilio
  • Weakness: less turnkey than Retell for non-developer teams

Bland

  • Cheapest at $0.09/min all-in
  • Optimized for high-volume outbound (dialing thousands of calls)
  • Less customization by design
  • Weakness: less LLM/prompt control

GPT-Live (once API lands)

  • Best conversational feel in the market
  • Reasoning delegation to GPT-5.5 built in
  • Weakness: no API yet as of July 12, 2026

The Bet for Q4 2026

Once GPT-Live API ships, expect ~30-40% of new voice-agent starts to build on it directly, another 30% to ship via Vapi/Retell/ElevenLabs orchestrators that pipe GPT-Live under the hood, and the rest to stay on their current stacks for lock-in reasons.

If you’re starting now: build on Retell (phone) or ElevenLabs (web), design the LLM layer to be swappable, and plan a Q1 2027 swap to GPT-Live if the API pricing comes in reasonably. Do not wait for GPT-Live API to start — you’ll lose two quarters of iteration.

Sources