OpenAI launched two voice products in late July 2026 that directly affect customer support teams: GPT-Live, a full-duplex voice model for desktop workers, and Presence, a managed enterprise platform for AI agents handling real support calls. Presence already powers OpenAI's own English-language phone support line at 1-888-GPT-0090, where the company says it resolves 75% of inbound issues without human assistance.
The enterprise voice rollout landed July 23, putting GPT-Live inside ChatGPT's desktop work surfaces. You can speak to ChatGPT while it starts tasks, checks progress, interrupts an agent, or coordinates several Work and Codex threads at once. That changes what voice does. It moves from a way to ask questions aloud to a way to steer work while keeping your hands free.
What full-duplex actually means for support calls
GPT-Live-1 and GPT-Live-1 mini launched July 8 as full-duplex voice models. The full-duplex distinction matters. The system can listen and speak at the same time instead of waiting for a clean pause before responding. You can interrupt it mid-sentence. It can keep listening while it delegates harder questions to another model in the background.
TechCrunch reported that OpenAI replaced its older Advanced Voice Mode with GPT-Live-1 mini by default for free users, while paid users gained access to the larger GPT-Live-1 model. The old voice flow worked like a turn-based meeting. This one handles interruption and background work, pauses included.
More than 150 million people use ChatGPT through Voice and Dictation each week, according to OpenAI's launch post. That base is now shifting to an architecture that treats voice as a continuous stream rather than a series of turns.
Presence is built for customer operations
Presence arrived separately on July 22. It is not self-serve ChatGPT Voice. OpenAI designed it as a managed enterprise product for voice and chat agents across customer support, billing issues, IT help desks, procurement, HR, claims, and sales workflows. Deployments go through Forward Deployed Engineers and select systems integrators, with access limited to eligible enterprise customers.
That service model signals who OpenAI is chasing. Salesforce has Agentforce. Microsoft has Copilot Studio. Contact-center AI startups have sold voice agents into support teams for years. OpenAI is now selling against all of them with a product that wraps models in policies, permissions, simulations, evaluations, guardrails, and escalation rules. It is less flashy than a new model name. It is also the part enterprises actually buy.
The strongest proof point is OpenAI's own support line. The company's Presence launch page says the system met or exceeded frontline human-support quality benchmarks within weeks and now resolves 75% of inbound issues without human assistance. The number is self-reported and hasn't been independently audited, but it comes from a real queue, not a demo clip.
OpenAI also said Presence's Codex-powered improvement loop reduced human handoffs by 15 percentage points in 10 days. A voice agent that works on launch day is useful. One that keeps adapting as policies, products, and customer behavior change is the harder promise. Customer operations break in the gaps. Refund rules change, account screens move, and suddenly the assistant that looked good in testing starts making mistakes.
The buyer still carries the operational burden
Presence does not remove the hard work from the company buying it. It makes that work visible. You still decide what an agent can see, what it can do, when it needs approval, and when a person takes over. OpenAI's product page names BBVA, SoftBank, and IAG as enterprises working with Presence, including BBVA's exploration of voice support for everyday banking in Mexico and SoftBank's testing of Japanese-language customer conversations. Those are precisely the places where vague automation promises meet regulated, high-pressure reality.
For professionals working in AI for Customer Support, the cost structure will shape adoption. Business workspaces include one hour of Voice in Chat, with extra use billed at five credits per minute. Voice in Work and Codex costs roughly six credits per connected minute where flexible pricing applies, while the underlying Work or Codex tasks still draw from the normal shared usage pool. Five or six credits per minute sounds manageable for a short test call. It looks different when a support team, IT desk, or engineering group leaves voice sessions open during real work.
Why this matters for customer support professionals
The open question is not whether voice AI can sound more natural. GPT-Live already moves that forward. The real question is whether companies will trust it with the messy work that touches customers and internal accounts, the stuff where errors have consequences. Presence is OpenAI's answer: bring engineers, policies, and guardrails, then prove it in production. For support leaders, the 75% self-reported resolution rate and the 15-percentage-point reduction in handoffs over 10 days are concrete claims to test against your own queue metrics. The product is enterprise software with all the boring obligations that come with it, and those obligations sit with the buyer, not the vendor.
Your membership also unlocks:
