AI agents · OpenClaw · self-hosting · automation

Quick Answer

GPT-Live-1 API vs GPT-Realtime-2.1 vs ElevenLabs Agents

Published:

The Short Answer

GPT-Live-1 (API)GPT-Realtime-2.1ElevenLabs Agents
Vendor / dateOpenAI · API since Sep 10, 2026 (ChatGPT since Jul 2026)OpenAI · Jul 6, 2026 (mini Jul 7)ElevenLabs · ongoing
ArchitectureFull-duplex: listens and speaks simultaneously; delegates reasoning to a backend modelSpeech-to-speech, turn-based, one model does voice + reasoningPlatform: STT/LLM/TTS orchestration with ElevenLabs voices, pluggable LLM
Voice-layer price$0.05/min$32 in / $64 out per M audio tokens ≈ $0.05/min; mini $10/$20 ≈ $0.016/min$0.08–$0.10/min reported, LLM bundled (~$0.08 on Business)
Reasoning cost+ backend model (Luna $0.20/$1.20, Terra $2/$12, Astra $10/$50 per MTok)IncludedIncluded (or your own LLM)
Interruptions / backchannelsNative — single model over both audio streamsTurn detection; handoffs on interruptionHandled by orchestration layer
TelephonyYesYesYes
Custom voicesExpanded library; custom voice via salesSmall setBest-in-class library + cloning
Best forTutoring, coaching, phone support where people talk over the agentStrict turn-based flows, one bill, cheap high volume (mini)Voice quality, brand voices, fast all-in-one build

Rule of thumb: if callers interrupt, GPT-Live-1. If you want one model and one line item, GPT-Realtime-2.1 (or mini). If the voice itself is the brand, ElevenLabs.

GPT-Live-1: what changed on September 10, 2026

OpenAI’s launch post makes one architectural claim and backs it with a number. The claim: chaining STT → LLM → TTS adds latency at every handoff and loses “timing, context, or the natural rhythm of a conversation.” The number: GPT-Live-1 improves Full Duplex Bench by 30 percentage points over GPT-Realtime-2.1, with the biggest gains in turn-taking latency and interactive behavior. Paired with GPT-6 Astra at medium effort it ranks #1 on Tau3, OpenAI’s end-to-end voice-agent task benchmark.

The API-specific additions:

  • Delegation. GPT-Live-1 handles the conversation while a backend text model (Astra, Terra, Luna, or a third-party model) does the reasoning and tool calls. OpenAI’s example: Luna for scheduling and order updates, Astra for complex support cases — you match reasoning depth and cost per task.
  • Steerable tone, pace and style through the system prompt.
  • Silence and background-noise handling without narrating every step aloud.
  • Long-session reliability, native ASR transcripts and response text, alphanumeric understanding, keyword biasing, and — although not turn-based — native turn detection for developers who still build around explicit boundaries.
  • Telephony support for full-duplex phone agents.
  • More voices across accents, dialects and languages; custom voices via sales.

Early customer data point: Speak reports GPT-Live-1 cut learner interruptions by almost 80% versus its previous turn-based tutor, because it gives people time to think before responding.

Pricing: $0.05 per minute for the voice layer. Add the backend model’s token cost. OpenAI Presence, its enterprise voice-agent product, now runs on GPT-Live-1.

GPT-Realtime-2.1: the turn-based baseline

Released July 6, 2026 (mini a day later), GPT-Realtime-2.1 is a single speech-to-speech model. Billing is per audio token — $32 per million in, $64 per million out, cached input $0.40 per million — which OpenAI’s rule of thumb (~600 tokens heard and ~1,200 spoken per minute) turns into roughly $0.05 per conversation minute. The mini variant at $10/$20 per million lands near $0.016 per minute, which makes it the cheapest OpenAI option for high-volume, scripted flows.

Strengths: one model, one bill, predictable turn boundaries, mature tooling. Weakness: it is the model GPT-Live-1 was benchmarked against on interruption handling, and it lost by 30 points.

ElevenLabs Agents: the voice-first platform

ElevenLabs sells the whole stack — its speech engine, an orchestration layer, telephony, and your choice of LLM — metered per conversation minute on top of a plan’s credit allotment. Third-party pricing guides in August–September 2026 put agent minutes at roughly $0.08–$0.10 (around $0.08 on Business tiers) after a 2026 price cut; confirm against your plan’s calculator, since minute rates sit on top of base credits and are the most common source of bill surprises.

What you buy is voice quality and breadth: the largest voice library, cloning, prosody control, and sub-second latency that is fast without being full-duplex. If the agent is the face of a consumer brand and the voice has to be yours, ElevenLabs still leads.

Cost worked example (10,000 minutes/month)

StackVoice layerReasoningApprox. total
GPT-Live-1 + Luna (light tasks)$500~$50–150 at Luna’s $0.20/$1.20 per MTok~$600
GPT-Live-1 + Astra (complex support)$500~$1,000–3,000 depending on tool calls and effort~$1,500–3,500
GPT-Realtime-2.1~$500included~$500
GPT-Realtime-2.1-mini~$160included~$160
ElevenLabs Agents~$800–1,000included~$800–1,000

Reasoning estimates assume a few thousand text tokens per minute of conversation; measure your own transcripts before committing. The pattern holds regardless: Realtime-mini is cheapest, GPT-Live-1’s cost is dominated by the backend you choose, and ElevenLabs charges a premium for voice.

Decision guide

SituationPick
Language tutoring, coaching, therapy-style conversationGPT-Live-1 + Terra or Astra
Inbound phone support with frequent interruptionsGPT-Live-1 + Luna/Terra, Astra for escalations
Outbound scripted calls at scaleGPT-Realtime-2.1-mini
Strict IVR-like turn boundaries, one vendor billGPT-Realtime-2.1
Branded voice, cloning, consumer app polishElevenLabs Agents
Already on ElevenLabs and happy with turn-takingStay; re-evaluate when full-duplex complaints appear

Sources