OpenAI Realtime (gpt-realtime-2.1) + GPT-Live-1
GPT-Live-1 parle, GPT-5.5 raisonne derrière. API au 10 sept. à 0,05 $/min pour la voix, plus le modèle. Scores The Batch contre AVM seulement.
Comparative Scores
Architecture
GPT-Live-1 plus GPT-5.5 is an architecture comparison: one model talks, another thinks. It is not a production French cascade until session cost and residency are remeasured. Vendor scores compare Live to Advanced Voice Mode only. GPT-Realtime-2’s public-arena ELO is 1079, rank 49, on 17 October 2026.
Analysis
The Batch (17 July 2026) describes ChatGPT Voice as two models: GPT-Live-1 or GPT-Live-1 mini speaks full duplex, and harder questions go to GPT-5.5 while the voice keeps talking. On GPT-Live-1, Instant calls GPT-5.5 Instant; Medium and High call GPT-5.5 Thinking. Mini only calls GPT-5.5 Instant. OpenAI’s own tests versus Advanced Voice Mode report GPQA 84.2% against 45.3%, BrowseComp 75.2% against 0.7%, and human preference 75.7% for GPT-Live-1 and 69.2% for mini. Those comparisons are not against other voices in this catalog. The 10 September API prices the voice layer at $0.05/min plus the model behind it. gpt-realtime-2.1 remains the dated speech-to-speech line and no Live-1 millisecond figure is published. Separately, GPT-Realtime-2 records an Artificial Analysis V2 Free ELO of 1077 on 5 October 2026; the Free tier does not publish a rank.
Strengths
- GPT-Realtime-2 : premier modèle vocal avec raisonnement GPT-5
- Contexte 128K (vs 32K pour GPT-Realtime-1.5)
- 5 niveaux de raisonnement : minimal, low, medium, high, xhigh
- Preambles : phrases de remplissage pendant le raisonnement
- Parallel tool calls audibles ("checking your calendar")
- GPT-Realtime-Translate : 70+ langues input → 13 output, $0.034/min
- GPT-Realtime-Whisper : STT streaming temps réel, $0.017/min
- ELO AA V2 Free 1077 (5 octobre ; rang non publié) — le rang 1208 de mai reste une annonce distincte
- EU Data Residency disponible
- +15.2% Big Bench Audio vs GPT-Realtime-1.5 (high)
- GPT-Live-1 : la voix parle, GPT-5.5 raisonne (Instant, Medium, High)
- GPQA 84,2 % et BrowseComp 75,2 % : chiffres OpenAI contre AVM, pas contre le catalogue
Weaknesses
- Pas de clonage vocal
- Pas de souveraineté (cloud US par défaut)
- GPT-Realtime-2 : $32/1M input + $64/1M output audio tokens (coûteux à l'échelle)
- 1.536s median latency (vs 0.864s Ultravox)
- Pas d'on-premise
- Aucune milliseconde publiée pour GPT-Live-1 ; le 1,536 s est le repère realtime antérieur
Voice Capabilities
No voice cloning. Pre-built voices uniquement.
Tone is steered through session instructions, with voice and speed selection.
How: `instructions`, `voice` and `speed` (0.25–1.5) on the Realtime session.
Validate: Tone instructions remain general: test a scene matrix rather than assume stable emotion.
WebSocket streaming. Full-duplex. Contexte 128K (vs 32K pour GPT-Realtime-1.5). Preambles (phrases de remplissage pendant le raisonnement). Parallel tool calls audibles. 5 niveaux de raisonnement : minimal, low (défaut), medium, high, xhigh. 1.536s median latency (vs 0.864s Ultravox).
No native lip-sync data.
Pricing
gpt-realtime-2.1 : $32/1M tokens audio entrée et $64/1M sortie, comme le palier 2 déjà noté. 2.1-mini : $10 / $20. GPT-Live-1 : $0,05/min pour la couche voix, plus GPT-5.5 derrière (Instant, ou Thinking en Medium/High). Translate $0,034/min et Whisper $0,017/min restent les tarifs de mai.
Sovereignty & Compliance
Cloud only. EU Data Residency disponible pour applications EU (Realtime API).
Data residency: US (default). EU Data Residency disponible (Realtime API).
OpenAI Realtime (gpt-realtime-2.1) + GPT-Live-1 — Strategic Positioning
Beyond technical specs: where does this tool sit in the ecosystem, what are the risks and strategic implications for GamiWays?
17 juillet 2026 : GPT-Live-1 parle, GPT-5.5 raisonne derrière (The Batch). Les scores GPQA et BrowseComp comparent Live à Advanced Voice Mode seulement. L’API du 10 septembre ajoute 0,05 $/min de voix, sans milliseconde. La ligne de mai (GPT-Realtime-2, Translate, Whisper) et l’ELO d’août restent la baseline datée.
A. Strategic Positioning
Target customer: Developer / Enterprise — GPT-5 reasoning voice agents, traduction live, STT streaming
7 mai 2026 : 3 nouveaux modèles audio. GPT-Realtime-2 (raisonnement GPT-5, contexte 128K, 5 niveaux de raisonnement, preambles, +15.2% Big Bench Audio). GPT-Realtime-Translate (70+ langues → 13 output, $0.034/min). GPT-Realtime-Whisper (STT streaming, $0.017/min).
B. Competitive Moat
- GPT-Realtime-2 : premier modèle vocal avec raisonnement GPT-5, contexte 128K, 5 niveaux de raisonnement
- Preambles : phrases de remplissage audibles pendant le raisonnement pour maintenir l'immersion
- GPT-Realtime-Translate : traduction live 70+ langues → 13 output, $0.034/min
- OpenAI brand and ecosystem — massive developer adoption and trust
Vulnerability: Cloud-only — no on-premise option. High cost. EU data sovereignty concerns. OpenAI's legal/regulatory exposure.
E. Strategic Questions for GamiWays
Sovereignty fit
Cloud-only with limited EU data residency. OpenAI's US jurisdiction creates sovereignty risk for Swiss/EU regulated deployments.
Build vs. Buy
Buy for Phase 1 if GPT-4o reasoning is required. For Phase 2 sovereignty, switch to open-source stack (Ultravox + Kokoro/Chatterbox).
Lock-in risk
Deep GPT-4o integration creates strong ecosystem lock-in. Switching costs are high if LLM reasoning is core to the voice agent design.
Roadmap alignment
Good for Phase 1 prototyping with GPT-4o. Incompatible with Phase 2 sovereignty requirements without major architectural changes.
This sheet's verification
The Batch, 17 Jul 2026 (GPQA, BrowseComp, preference vs AVM). API price: OpenAI, 10 Sep 2026. Public Speech Arena, 17 Oct 2026: GPT-Realtime-2 ELO 1079, rank 49.
Update note: Arena publique du 17 octobre 2026 : GPT-Realtime-2 ELO 1079, rang 49. Pas de nouveau snapshot API : clé absente.
This date covers provider-specific prices, capabilities and notes. Comparative benchmarks follow the synchronization and methodology shown above.
ELO benchmarks / indices: Artificial Analysis · Last API sync : 5 October 2026