<p><strong>CONFIRMED.</strong> OpenAI released GPT-Live-1 in the API on September 10, 2026. The full duplex voice model arrives with telephony support, native transcription, response text, and a listed price of $0.05 per minute for the voice layer. That is the difference between watching a convincing demonstration and putting the technology inside a product.</p><h2>The API has finally arrived</h2><p>GPT-Live-1 was already powering ChatGPT’s new voice experience, but access for application developers remained a promise. OpenAI now says the model is available in the API and lets each team choose the text model, tools, and agent harness running behind the conversation. The voice layer handles continuous interaction, while heavier work can be delegated to a reasoning model such as GPT-6 Astra or to a third party model.</p><h2>A voice layer that does not have to stop to think</h2><p>The advantage is not simply a better sounding voice. GPT-Live-1 listens and speaks at the same time, instead of waiting for a clean silence window before deciding that the user is finished. That improves interruptions, pauses, short acknowledgements, and changes of direction. OpenAI also says the model handles silence, background noise, and long sessions better. In an early evaluation, Speak recorded almost 80% fewer tutor interruptions compared with earlier turn based systems.</p><p>The company also reports a 30 percentage point improvement on Full Duplex Bench against GPT-Realtime-2.1. Those figures come from OpenAI, not from an independent benchmark, but they point to the part of voice that matters most: conversational timing, not transcription quality alone.</p><h2>What changes for Brazilian businesses</h2><p>For small and midsize companies, the release creates a shorter path to phone support, reservation confirmation, order updates, support triage, and information collection. A restaurant can answer calls without assembling separate speech recognition, text model, and speech synthesis layers. A clinic can let an agent keep the conversation moving while it checks schedules and internal systems. An events operation can handle questions and changes without turning every customer pause into a turn detection error.</p><p>The $0.05 per minute price applies only to the front end voice layer. The backend model, tools, telephony, infrastructure, observability, and possible searches add to the bill. The headline price is therefore an entry point, not the total cost of a call. For Brazil, exchange rates, taxes, Portuguese support, and performance with proper names must be measured before any migration.</p><h2>The caveat that matters</h2><p>Full duplex does not remove product responsibilities. Developers still own permissions, policies, human handoffs, call records, and cost controls. The freedom to choose the backend is powerful, but it also means the final experience can vary widely between implementations. In addition, proprietary evaluations do not prove superior performance in Brazilian Portuguese, call center noise, or hostile customer service scenarios.</p><h2>Our editorial read</h2><p>This launch matters because it narrows the distance between natural voice and useful automation. Our read is that OpenAI is not selling only a speech model. It is trying to turn conversational timing into an interface layer while keeping deeper intelligence modular. Builders in Brazil should test it now on a narrow workflow and measure interruptions, latency, total cost, and human escalation. Rebuilding an entire architecture based on the announcement would be marketing. Ignoring the shift would be a mistake too.</p><h2>Sources</h2><p>OpenAI, API announcement, September 10, 2026:<br />https://openai.com/index/introducing-gpt-live-1-in-the-api<br /><br />OpenAI, GPT-Live systems engineering, August 3, 2026:<br />https://openai.com/index/continuous-voice-interaction-with-gpt-live/<br /><br />VentureBeat, original GPT-Live launch, July 8, 2026:<br />https://venturebeat.com/technology/openai-launches-gpt-live-a-full-duplex-voice-upgrade-that-lets-chatgpt-talk-more-like-a-person</p>