Name: GPT-Live-1
Author: OpenAI

OpenAI launched GPT-Live-1 and GPT-Live-1 mini on July 8, 2026, replacing Advanced Voice Mode in ChatGPT with a pair of full-duplex voice models that can listen and speak at the same time. The key architectural shift: instead of chaining separate speech-to-text, language model, and text-to-speech components, GPT-Live processes audio continuously and makes decisions many times per second about whether to speak, keep listening, pause, or call a background tool.

TL;DR

Full-duplex voice model that listens and speaks simultaneously - no more waiting for silence before responding
GPQA accuracy jumps from 45.3% (Advanced Voice Mode) to 84.2% (GPT-Live-1 High); BrowseComp goes from 0.7% to 75.2%
GPT-Live-1 mini is free in ChatGPT; GPT-Live-1 for paid tiers; API access planned but not yet live

The model family replaces what OpenAI called "Advanced Voice Mode," which processed audio in discrete turns and waited for the user to stop talking before responding. GPT-Live-1 is the flagship version for Go, Plus, and Pro subscribers. GPT-Live-1 mini is the default for Free users and is notably more capable than the voice experience it replaces. Both models are rolling out globally across iOS, Android, CarPlay, and ChatGPT.com.

The product lead for GPT-Live is Atty Eleti. OpenAI reported that more than 150 million people use ChatGPT voice features each week - making this one of the widest-surface AI model releases the company has done.

Key Specifications

Specification	Details
Provider	OpenAI
Model Family	GPT-Live
Parameters	Not disclosed
Context Window	Not disclosed
Input Price	Not yet available (API forthcoming)
Output Price	Not yet available (API forthcoming)
Release Date	July 8, 2026
License	Proprietary

Benchmark Performance

The most striking numbers from the launch are the gaps between GPT-Live-1 and its predecessor on standard LLM benchmarks, made possible by the background delegation to GPT-5.5:

Benchmark	GPT-Live-1 (High)	Advanced Voice Mode	Notes
GPQA	84.2%	45.3%	Expert science reasoning
BrowseComp	75.2%	0.7%	Agentic web search
t3-Voice Telecom	~65% task completion	~30% task completion	Multi-turn support tasks

The GPQA jump from 45.3% to 84.2% is the headline number - that's almost a 2x improvement on a benchmark that tests PhD-level reasoning in biology, chemistry, and physics. The BrowseComp figure is even more striking: Advanced Voice Mode was effectively unable to complete agentic web search tasks (0.7%), while GPT-Live-1 High reaches 75.2%. Both gains reflect the delegation architecture rather than the voice model itself becoming a stronger reasoner - GPT-Live-1 offloads hard queries to GPT-5.5 Thinking in the background.

On the internal t3-Voice Telecom benchmark, which measures completion of realistic telecom customer support tasks in a multi-turn setting, GPT-Live-1 High completes roughly 65% of tasks at around 385 seconds versus Advanced Voice Mode's 30% completion rate. That's the kind of benchmark that maps to real enterprise voice agent use cases, and the improvement is meaningful.

It's worth being skeptical of the first two benchmarks: GPQA and BrowseComp aren't designed for voice models and the scores reflect the backend GPT-5.5 reasoning rather than voice-specific quality. OpenAI also reported strong preference in human evaluations of conversational flow and pleasantness, but those studies are conducted by the company and the methodology isn't independently verified.

GPT-Live-1 voice interface in ChatGPT showing the new full-duplex voice mode The new GPT-Live voice interface in ChatGPT, showing the simultaneous listen-and-speak capability rolling out across iOS, Android, and web. Source: 9to5mac.com

Key Capabilities

Full-Duplex Architecture

The core capability is simultaneous input and output processing. Previous voice systems worked in half-duplex: the model waited for a speaker to pause, transcribed the audio, ran the language model, synthesized speech, then played it back. GPT-Live runs all of these continuously in parallel. The result is that the model can be interrupted mid-sentence (it'll actually listen and adjust), can produce backchannel responses like "mhmm" or "yeah" while you're still speaking, and can handle pauses without interpreting them as turn-endings. Users can explicitly tell it to stay silent - saying something like "just hold on" keeps it from jumping in.

Three Reasoning Variants

GPT-Live-1 ships with three reasoning configurations that select the backend model:

GPT-Live-1 Instant - Uses GPT-5.5 Instant as the background model. Lowest latency, best for quick-fire conversation.
GPT-Live-1 Medium - Uses GPT-5.5 Thinking at medium reasoning effort. Better at multi-step queries without long waits.
GPT-Live-1 High - Uses GPT-5.5 Thinking at high reasoning effort. Where the top benchmark numbers come from. Slower but substantially more accurate on hard tasks.

GPT-Live-1 mini uses GPT-5.5 Instant as its backend and is the default for free users. You can see how GPT-5.5 Instant and GPT-5.5 stack up as base models in their own pages.

Natural Conversation Behaviors

OpenAI has specifically engineered for behaviors that make voice feel less mechanical: the model adjusts speaking pace when asked, handles overlapping speech (you can talk over it), and can remain silent for extended periods - useful in listening contexts. During internal testing, conversations of 30-40 minutes were sustained without degradation. Live translation is also supported via the full-duplex timing - the model can translate as it listens rather than waiting for sentence completion.

The model also surfaces visual cards for structured data - sports scores, weather, stock prices - on-screen while the audio continues. This is a narrower feature than screen sharing or video input, which aren't available at launch.

Pricing and Availability

As of July 8, 2026, GPT-Live-1 is available in ChatGPT only:

Free tier: GPT-Live-1 mini, replacing the previous Advanced Voice Mode
Go, Plus, Pro tiers: GPT-Live-1 (full model)

API access is planned. OpenAI has a developer signup form at openai.com/form/gpt-live-1-in-the-api for notification when it becomes available. No pricing has been announced for the API.

For context, the previous generation GPT-Realtime-2 was priced at $4.00/M input tokens, $16.00/M output text tokens, and $64.00/M audio output tokens in the API. GPT-Live-1 pricing will likely be disclosed when API access opens.

The rollout is global across iOS, Android, CarPlay (via the ChatGPT app integration introduced in iOS 26.4), and ChatGPT.com.

Strengths and Weaknesses

Strengths

Full-duplex architecture eliminates the half-duplex turn-taking of every previous ChatGPT voice mode
Backend delegation to GPT-5.5 delivers near-frontier reasoning performance from within a voice interface - see our AI voice speech leaderboard for broader context
GPQA and BrowseComp gains over predecessor are major and independently meaningful benchmarks
Free tier gets GPT-Live-1 mini, a meaningful upgrade from what was there before
Safety guardrails run in real time: can steer responses, surface crisis resources, or end the conversation in high-risk situations
Nine remastered voices with protections against unauthorized voice imitation

Weaknesses

API not yet available at launch - developers cannot build on it until the waitlist opens
No video or screen sharing at launch, which limits visual context use cases that Google Gemini Live covers
Multilingual parity is uneven: OpenAI acknowledged that some languages get non-native accents. The Hindi demo at launch showed a "heavy American accent" with an "unnatural sounding, slightly bookish tone"
Benchmark gains (GPQA, BrowseComp) reflect GPT-5.5 backend reasoning, not voice model capability - the voice layer itself isn't what's scoring 84.2% on GPQA
Context window, parameter count, and latency targets aren't publicly disclosed
No third-party independent evaluation available yet - all performance data comes from OpenAI's own testing

GPT-Realtime-2 - The predecessor API voice model that GPT-Live-1 supersedes in ChatGPT
GPT-5.5 - The frontier model powering GPT-Live-1's background reasoning
GPT-5.5 Instant - The low-latency backend used by GPT-Live-1 Instant and GPT-Live-1 mini
AI Voice Speech Leaderboard - How voice models compare across providers

FAQ

What is the difference between GPT-Live-1 and GPT-Live-1 mini?

GPT-Live-1 is for paid ChatGPT subscribers (Go, Plus, Pro) and uses GPT-5.5 Thinking at Instant, Medium, or High reasoning levels. GPT-Live-1 mini is the free-tier default and uses GPT-5.5 Instant only.

No. Video and screen sharing are not available at launch. The model supports visual cards for structured data (weather, sports, stocks) but cannot see your screen.

When will GPT-Live-1 be available via API?

OpenAI has not announced a date. Developers can sign up for notification at openai.com/form/gpt-live-1-in-the-api.

How does GPT-Live-1 handle voice imitation safety?

The model uses a set of nine approved voices remastered for the GPT-Live system and includes protections against unauthorized imitation of real people's voices. Real-time safety steering can end conversations or surface resources in high-risk situations.

What languages does GPT-Live-1 support?

OpenAI states the model is optimized for "most popular languages" without a complete list. The company acknowledged that some languages may have non-native accents or fluency gaps. Hindi was cited as an example with accent issues at launch.

Sources: