GPT-Live-1.1 improves voice conversations, but reveals security regressions

OpenAI is rolling out GPT-Live-1.1 and Quip in ChatGPT Voice, featuring new voices and improved accents. Their evaluation also reveals security limitations.

Talking to an AI without waiting for a clear turn to end, interrupting it mid-response, or pausing before continuing a thought. With GPT-Live-1.1 and GPT-Live-1.1 Quip, OpenAI is updating the models behind ChatGPT's real-time voice conversations. The release introduces new voices and improvements to accents across several languages, including French. However, the accompanying system card presents mixed safety results: while both models remain broadly comparable to their predecessors on evaluations drawn from real user interactions, several statistically significant regressions appear in synthetic safety tests. Two Models for Real-Time Conversations

GPT-Live-1.1 succeeds GPT-Live-1, while GPT-Live-1.1 Quip replaces GPT-Live-1 mini.

Both models use a full-duplex architecture, allowing them to listen and respond continuously without requiring a strict alternation between speaking and listening.

This enables the system to account for pauses, interruptions, and changes in conversational pace.

A user might begin a sentence, hesitate, or interrupt the assistant. The model must determine whether to respond, continue speaking, or keep listening.

These capabilities were already part of the previous GPT-Live generation. The 1.1 versions build on that foundation rather than introducing an entirely new architecture.

According to OpenAI's official documentation, GPT-Live-1.1 becomes the default voice model for Plus and higher-tier users, including supported voice experiences in Codex and Dots. Quip serves other users. New Voices and More Natural Accents

The rollout also expands ChatGPT's available voices.

OpenAI reports improved accent rendering in several languages, including French, German, Hindi, Japanese, Korean, and Mexican Spanish.

Additional voice options are being introduced for those languages, alongside Bubbie in English.

The changes are intended to better reflect pronunciation and vocal characteristics across languages during conversations.

The assistant also uses a broader selection of short phrases to indicate when it is searching, reducing repetition during those moments.

The ChatGPT release notes specify that available voices may vary by region and account.

Voice selection is accessible through ChatGPT's Voice settings.

Public release information also points to improvements in instruction steering and delegation to other models. The system card does not, however, provide a detailed comparative metric quantifying these conversational improvements. Broadly Stable Results on Real-World Voice Prompts

OpenAI reused evaluations originally developed for the first GPT-Live release.

The first set uses audio examples from users who agreed to share their voice interactions to improve the models.

These examples undergo privacy and eligibility checks, including steps to reduce personal information, before being included in testing.

The evaluation covers six safety categories: sexual content, illicit behavior, mental health, personal data, emotional reliance, and self-harm.

For sexual content, GPT-Live-1.1 scores 0.81 compared with 0.77 for GPT-Live-1. Quip improves from 0.69 to 0.78 relative to GPT-Live-1 mini.

On mental health, both new models score 0.81, compared with 0.78 and 0.76 for their respective predecessors.

Personal data protection remains at 0.92 for GPT-Live-1.1, while Quip rises from 0.89 to 0.92.

Other results change only slightly. Illicit behavior performance moves from 0.87 to 0.86 for the larger model, while emotional reliance remains at 0.76.

OpenAI reports no statistically significant regressions across this evaluation set.

However, the examples were deliberately selected for difficulty. Their scores do not represent the frequency of safety failures across all production voice conversations. Significant Regressions in Synthetic Safety Tests

A second evaluation uses synthetically generated audio prompts designed to target difficult situations, safety boundaries, and scenarios that may rarely appear in ordinary interactions.

The results are less consistent.

For GPT-Live-1.1, the mental health score falls from 0.70 to 0.63, while the gore category drops from 0.76 to 0.69.

Other declines appear in sexual content, from 0.87 to 0.84, and emotional reliance, from 0.86 to 0.81.

OpenAI identifies these four changes as statistically significant regressions relative to GPT-Live-1.

Quip also scores below GPT-Live-1 mini in certain categories. Its mental health score changes from 0.60 to 0.59, while the gore score falls from 0.66 to 0.62.

Other results remain comparatively strong, including personal data protection, which reaches 0.95 for GPT-Live-1.1.

OpenAI says further research is needed to understand the declines and improve future versions.

These evaluations focus on the models themselves and do not account for every additional safety layer used in products, including crisis-related interventions and parental controls. Why the Evaluation Methods Produce Different Results

The differences between stable real-world prompt results and regressions in synthetic testing partly reflect how the two datasets are constructed.

Real-world interaction tests often involve ambiguous requests, longer conversation histories, imprecise wording, and repeated attempts to obtain problematic responses.

Synthetic tests target clearly defined safety scenarios more directly, including rare situations that may be difficult to collect from user conversations.

OpenAI treats the methods as complementary.

The company also cautions that scores from separate system cards are not necessarily directly comparable, because reference models, evaluation methods, datasets, and graders can change over time.

The GPT-Live-1.1 findings should therefore be interpreted within the comparisons conducted for this particular release. How Models Respond to Restrictions

Beyond conversational safety, OpenAI examined how the models behave when encountering restrictions during computer-based tasks.

These evaluations adapt alignment tests originally developed for non-voice models.

On