GPT-Live (GPT-Live-1 / GPT-Live-1 mini)

GPT-Live is a family of voice models developed by OpenAI, launched on July 8, 2026 in two versions — GPT-Live-1 and GPT-Live-1 mini — that replaced Advanced Voice Mode as the engine of ChatGPT Voice. The models use a full-duplex architecture that listens and speaks simultaneously rather than in discrete turns, and delegate questions requiring search, reasoning, or agentic work to a background frontier model, GPT-5.5 at launch (Source: openai.com).

Architecture

OpenAI describes GPT-Live as addressing the limitations of two prior generations of voice systems: cascaded pipelines (the original ChatGPT Voice chained speech-to-text, a language model, and text-to-speech, losing information between models) and turn-based voice models (Advanced Voice Mode processed audio in a single model but required silence-based turn detection, causing interruptions at unnatural times) (Source: openai.com).

The family makes two architectural changes. First, continuous interaction: the model processes input while generating output and makes interaction decisions many times per second — whether to speak, keep listening, pause, interrupt, or invoke a tool — enabling back-channel responses ("mhmm"), live translation, and a persistent sense of time. Second, delegation: GPT-Live decouples conversational interaction from deeper work, handing tasks requiring web search, reasoning, or agentic capabilities to another model (GPT-5.5 at launch) while maintaining the conversation, with OpenAI stating it will continuously update the delegated model as new frontier models release (Source: openai.com).

Capabilities and benchmarks

OpenAI reports that GPT-Live-1 and GPT-Live-1 mini are strongly preferred over Advanced Voice Mode in matched 5–10 minute human-evaluation conversations measuring overall preference, turn-taking, interruptions, conversational flow, and naturalness. On capability benchmarks, OpenAI reports that GPT-Live-1 substantially outperforms Advanced Voice Mode on GPQA (expert-level scientific reasoning), shows strong gains on BrowseComp (agentic web search), and outperforms it on an internal τ³-Voice Telecom variant testing multi-turn voice support tasks (Source: openai.com). Press coverage of the launch reported a GPQA score of 84.2% at high reasoning versus 45.3% for Advanced Voice Mode, and more than 150 million weekly ChatGPT voice users (Source: cnbc.com).

Safety and evaluations

The GPT-Live system card reports voice-native safety evaluations built from production prompts and synthetically generated audio prompts targeting policy edge cases, across categories including sexual content, illicit behavior, mental health, personal data, emotional reliance, self-harm, hate, and gore. OpenAI reports the models perform comparably to or better than Advanced Voice Mode across nearly all evaluated areas, with a slight, not statistically significant regression for GPT-Live-1 on emotional reliance (0.88 → 0.82 on production prompts) (GPT-Live System Card (OpenAI, July 2026)). The card distinguishes the two evaluation sets by what each can establish: strong synthetic-set performance "indicates that safety training is transferring for clear, intentionally constructed risks, but may not translate to production behavior," while the production set carries "more ambiguous or borderline context, longer interaction histories, and persistent attempts to steer the model toward unsafe outputs." Neither set is prevalence-weighted; both are built around cases where the Advanced Voice Mode models were not yet responding ideally.

Because voice conversations unfold in real time, the system adds runtime safeguards that can steer the model mid-response, surface safety messaging or resources, play a spoken safety message, or end the conversation in higher-risk cases; self-harm conversations adapt ChatGPT's support flows for voice, including crisis-helpline support. Teen protections include age-appropriate behavior trained into the model, parental controls over voice access, and notification of linked parents in higher-risk situations involving signs of potential self-harm. OpenAI states the models use predefined voices with safeguards against imitating a real person's voice, and that post-launch monitoring will focus on emotional reliance (Source: openai.com; GPT-Live System Card (OpenAI, July 2026)). Red teaming ran in three rounds — a baseline before any safety training, then two validation rounds — across child-coded voice, impersonation, speaker identification, sensitive-trait identification, self-harm, emotional reliance, scams and manipulation, and audio-specific perturbations; early findings directed mitigation effort to sexual content, emotional reliance, and self-harm, while voice cloning and impersonation were found "policy-compliant by default" (GPT-Live System Card (OpenAI, July 2026)).

Under the Preparedness Framework, OpenAI's Safety Advisory Group determined that neither model, operating without delegation, could plausibly be considered High in any Tracked Category (Biological and Chemical Risk, AI Self-Improvement, Cybersecurity); delegated work inherits the safeguards of the underlying model, cybersecurity risk is described as constrained at launch because the models lack independent tool access and code execution, and AI Self-Improvement capability evaluations were not run because the models are less capable than GPT-5.5 Thinking on intelligence evaluations. OpenAI states it "will reassess the cybersecurity safeguards posture before enabling additional tools" (GPT-Live System Card (OpenAI, July 2026)).

Availability and limitations

The models rolled out globally on July 8, 2026 across iOS, Android, and chatgpt.com, with GPT-Live-1 becoming the default for Go, Plus, and Pro users and GPT-Live-1 mini for Free users; API access is planned. Reasoning depth is user-selectable (Instant, Medium, High, backed by GPT-5.5 Instant or Thinking). At launch GPT-Live does not support voice with video or screen sharing, and OpenAI states some languages may show non-native accents or fluency gaps; legacy Standard and Advanced Voice Mode remain available where those features are needed. The ChatGPT Voice experience adds visual cards (weather, stocks, sports, maps) rendered during conversation (Source: openai.com).

Relationships

• depends-on: GPT-5.5 ('Spud') — delegated background model at launch
• instance-of: AI Companions — voice-interface interaction layer; emotional-reliance monitoring
• related: GPT-5 Family (OpenAI), GPT-5.6 (Sol, Terra, Luna), OpenAI, OpenAI Preparedness Framework V.2