OpenAI Opens GPT-Live-1 Voice API to Developers at 5 Cents Per Minute

30-Second Summary

OpenAI made its GPT-Live-1 voice model available to developers via API on 13 September 2026 at five cents per minute for the voice layer. The model handles full-duplex conversation, meaning it listens and speaks simultaneously without the pause-and-respond pattern that made earlier voice AI sound mechanical. In standardised testing across airline, retail, and telecom support scenarios, GPT-Live-1 resolved 83.6 per cent of tasks on the first attempt. The previous OpenAI voice model completed 45.7 per cent on the same benchmark. For business operators running lean teams, this is the first commercially priced voice agent capable of handling routine customer calls without human supervision at a cost that holds up under scrutiny.

At a Glance

• Topic: Agent Systems
• Company: OpenAI
• Date: 13 September 2026
• Announcement: GPT-Live-1 API opened to developers at $0.05 per minute
• What Changed: Voice AI is now full-duplex, benchmarked on real business scenarios, and commercially priced for non-enterprise deployments
• Why It Matters: The cost and quality threshold for AI phone agents has dropped into range for businesses with 10 to 200 employees
• Who Should Care: Any operator running phone-based customer support, scheduling, or reservation workflows

Key Facts

• Company: OpenAI
• Launch Date: 13 September 2026 (API developer access; ChatGPT users received GPT-Live in July 2026)
• What Changed: Full-duplex voice model available via API, replacing the turn-based text pipeline architecture of earlier voice models
• Who It Affects: Businesses running phone support, inbound enquiries, scheduling, or reservations
• Primary Source: OpenAI (openai.com/index/introducing-gpt-live)

What Happened

OpenAI opened its GPT-Live-1 voice model to developers on 13 September 2026, making full-duplex voice AI commercially available via API at $0.05 per minute. The model was introduced to ChatGPT users on 9 July 2026 and has now been made accessible to businesses and developers building their own voice-enabled products and services.

GPT-Live-1 uses a full-duplex architecture, which means the model processes incoming audio and generates speech simultaneously. This removes the turn-taking pattern that made earlier voice AI models feel mechanical and created jarring delays during calls. Callers can interrupt, change direction, or add context mid-sentence without breaking the conversation flow.

OpenAI measured the model's performance using Tau3, a standardised benchmark covering airline, retail, and telecom customer support scenarios. Paired with GPT-6 Astra at medium reasoning effort, GPT-Live-1 completed 83.6 per cent of Tau3 tasks on the first attempt. The previous benchmark holder, GPT-Realtime-2.1, completed 45.7 per cent on the same test. OpenAI also reports a 30-point gain on Full Duplex Bench over GPT-Realtime-2.1. The model launches with 12 voices across accents and languages; custom voice arrangements require a separate agreement through OpenAI's sales team.

The $0.05 per minute price covers the voice layer only. The underlying reasoning model, whether GPT-6 Astra or another supported option, is billed separately at standard token rates. Total per-call cost depends on call duration and the complexity of reasoning the task requires.

Why It Matters

• Resolution rate changed the equation. A voice agent that resolves 83 per cent of routine calls on the first attempt is no longer a novelty demonstration. It is a viable first line of support for businesses that currently rely on staff for high-volume, low-complexity calls.
• Full-duplex removes the main reason customers dislike voice bots. The mechanical pause-and-respond pattern is gone. Callers experience a conversation, not a scripted menu system.
• Five cents per minute is within reach for small businesses. A ten-minute call costs $0.50 in voice fees plus model costs. For businesses receiving hundreds of routine calls per month, the economics are straightforward to model and compare against current staffing costs.
• The API targets use cases lean teams currently rely on staff for. Phone support, scheduling, reservations, and inbound enquiries are exactly where GPT-Live-1 is positioned and benchmarked.
• Pilots do not require an enterprise contract. Developers and technical operators can access the API directly, making a scoped test feasible without a lengthy procurement process.

The David and Goliath View

The reason most businesses with 10 to 200 employees have not deployed voice AI is not that they lacked awareness. They lacked a product that actually worked at a price that made sense. Voice AI has historically failed on at least one of two grounds: the model was too expensive for sustained deployment, or the call quality was poor enough to damage customer relationships. GPT-Live-1 addresses both problems in the same release.

The 83.6 per cent Tau3 resolution rate matters not because it is perfect, but because it crosses the practical threshold where an automated agent saves more time than it creates in escalations and complaints. The 16 per cent of calls that do not resolve on the first attempt still need human follow-up, but a system that automates 84 per cent of routine call volume frees the team to focus on the cases where human judgement genuinely adds value. That shift, from humans handling everything to humans handling the 16 per cent that needs them, is where lean organisations build compounding efficiency.

The recommended move is a narrowly scoped pilot. Identify the three to five call types your team handles on repeat every week. Confirm they resemble the Tau3 task categories: structured enquiries, scheduling requests, account queries, and reservation changes. Run a controlled test over 30 days, measure first-contact resolution against your current baseline, and make a deployment decision based on your own data rather than benchmark figures alone.

Where This Fits in the AI Stack

AI Growth Engine: Voice agents handling inbound enquiries, qualification calls, and scheduling allow sales and service capacity to grow without proportional headcount increases.

Employee Amplification Systems: GPT-Live-1 automates the high-volume, low-complexity call types that currently consume frontline staff time, freeing those staff for higher-value customer interactions that require human judgment.

Questions Operators Are Asking

Can I deploy this without a technical team? The current GPT-Live-1 release is an API for developers, not a no-code product. Deploying it requires either a developer or a third-party voice AI platform that has integrated the GPT-Live-1 API. Several voice AI platforms are expected to build GPT-Live-1 into their products in the weeks following the API launch. If your business does not have developer resources, the practical path is to identify which platforms in your existing stack are building on GPT-Live-1 and wait for their native integration.

What happens to calls the agent cannot handle? GPT-Live-1 is designed to escalate calls it cannot resolve. OpenAI's documentation indicates that escalation logic is configurable, allowing businesses to route unresolved calls to a human agent or a callback queue. The 16.4 per cent of Tau3 calls not resolved on the first attempt did not all result in failed interactions; some completed on a second attempt or via a different resolution path.

Is customer data safe during a call? OpenAI's API terms govern how audio input and conversation data are handled during a GPT-Live-1 call. Operators should review the current API data usage policy before deployment and confirm data residency and retention settings for any workflow involving customer account details or payment information.

What does a realistic pilot cost? A 500-call pilot with a seven-minute average call length would cost roughly $175 in voice layer fees ($0.05 multiplied by 7 minutes multiplied by 500 calls), plus model inference costs on top. This is a manageable test budget for most businesses and sufficient volume to generate statistically useful resolution rate data.

Will this replace our customer support team? GPT-Live-1 is best suited to the repetitive, structured enquiries that make up the routine volume of most support queues. Complex queries, complaints requiring judgement, and relationship-sensitive conversations still need human handling. The practical outcome for most businesses is fewer after-hours missed calls and shorter response times on high-volume routine enquiries, not an immediate reduction in support headcount.

Citable Summary

What happened: OpenAI opened the GPT-Live-1 full-duplex voice API to developers on 13 September 2026 at $0.05 per minute, recording an 83.6 per cent first-attempt resolution rate on standardised customer service benchmarks.

Why it matters: The combination of full-duplex architecture, commercial pricing, and verified resolution performance makes AI phone agents economically viable for businesses with 10 to 200 employees for the first time.

David and Goliath view: Operators should run a narrowly scoped 30-day pilot on their most repetitive call types and measure their own resolution rate before redesigning their support model around GPT-Live-1.

Offer relevance:

• AI Growth Engine: Inbound enquiry, qualification, and scheduling at scale without proportional headcount growth
• Employee Amplification Systems: Frontline staff freed from routine call volume to focus on higher-value customer interactions