Voice Models OpenAI Launched GPT-Live on July 8. Developers Are Still on the Waitlist. Published 26 July 2026 | Last updated 26 July 2026 Verified by Grok 26 July 2026 • AI Race Facts verification protocol OpenAI launched GPT-Live on 8 July 2026 with a new full-duplex voice architecture for ChatGPT. Consumers received it immediately. As of 25 July — 17 days later — the model is still not available in the OpenAI API. The system card released with the launch also reveals that one safety category performed worse on real user prompts than the synthetic results emphasized at launch. TL;DR GPT-Live launched in ChatGPT Voice on 8 July 2026 with full-duplex audio. As of 25 July 2026, it remains unavailable in the OpenAI API. On production prompts, emotional reliance safety scores fell from 0.88 to 0.82 for GPT-Live-1. The same category improved on synthetic prompts (0.72 → 0.91). Preparedness Framework clearance applies when operating without delegation . What consumers received on July 8 GPT-Live-1 and GPT-Live-1 mini introduced full-duplex voice in ChatGPT, allowing the model to listen and speak at the same time with more natural interruptions. It rolled out to paid and free users in supported regions on chatgpt.com and the mobile apps. OpenAI described it as a significant improvement in conversational feel over Advanced Voice Mode. What developers still do not have CHECKED-PRIMARY Seventeen days after launch, GPT-Live remains unavailable in the OpenAI API. Developers continue to see a waitlist notification. This creates a clear gap between the experience available to ChatGPT users and what developers can access programmatically. The safety finding the launch post did not highlight The launch post stated GPT-Live performed “comparably to or better than Advanced Voice Mode across nearly all of the areas we evaluated.” The system card makes clear what that phrasing leaves out. CLAIM On the production-prompt evaluation using real user audio, emotional reliance safety scores for GPT-Live-1 fell from 0.88 to 0.82. On synthetic prompts, the same category improved from 0.72 to 0.91. DERIVED This represents a 0.06 decline on real production prompts and a 0.19 gain on synthetic prompts in the same risk category. OpenAI notes in the card that strong synthetic performance shows safety training transfers well for clearly constructed risks, but production prompts involve more ambiguous context, longer histories, and persistent steering attempts. The one category that declined on production data is the type of real-world risk the card flags as potentially different from synthetic testing. OpenAI states that neither regression (emotional reliance for GPT-Live-1 and sexual content for the mini model) is statistically significant. The evaluations are not prevalence-weighted and were built around difficult cases where previous models already underperformed. They are not presented as real-world safety rates. The delegation caveat in the system card The Preparedness Framework section clears GPT-Live for biological, chemical, and cybersecurity risk only when operating **without delegation**. When the model delegates work to another model, the safety posture inherits the safeguards of that underlying model. AI Self-Improvement evaluations were not run, as GPT-Live is considered less capable than GPT-5.5 Thinking in that area. Sources OpenAI, "Introducing GPT-Live" , 8 July 2026. OpenAI, "GPT-Live System Card" , July 2026. OpenAI Help Center, ChatGPT Release Notes, July 2026 entries. Published 26 July 2026 | Last updated 26 July 2026 • Verified by Grok