Lab publication Credible

GPT-Live-1.1 system card

OpenAI's system card for GPT-Live-1.1 and GPT-Live-1.1 Quip, full-duplex voice models that succeed GPT-Live-1 and GPT-Live-1 mini as ChatGPT's default voice models. It reports voice-native disallowed-content evaluations on de-identified production audio and on synthetic audio prompts, audio versions of alignment evaluations (restriction circumvention, deception), and a Preparedness Framework statement. On the synthetic set, OpenAI reports statistically significant regressions for GPT-Live-1.1 relative to GPT-Live-1 in the sexual, mental health, emotional reliance and gore categories.

Read at deploymentsafety.openai.com Archived copy

Publisher

OpenAI

Published

15 Oct 2026

Added

3 days ago

DOI

—

Key Findings

Synthetic-prompt voice evaluations (Table 2; columns GPT-Live-1 / GPT-Live-1.1 / GPT-Live-1 mini / GPT-Live-1.1 Quip): mental health 0.70 / 0.63 / 0.60 / 0.59; emotional reliance 0.86 / 0.81 / 0.80 / 0.80; self-harm 0.89 / 0.87 / 0.85 / 0.86; gore 0.76 / 0.69 / 0.66 / 0.62.
OpenAI states that GPT-Live-1.1 has statistically significant regressions on sexual, mental health, emotional reliance and gore categories compared with GPT-Live-1, that more research is needed to understand them, and that it plans to improve these categories in future releases.
Production-prompt voice evaluations (Table 1, same columns): mental health 0.78 / 0.81 / 0.76 / 0.81; emotional reliance 0.76 / 0.76 / 0.70 / 0.79; self-harm 0.85 / 0.86 / 0.84 / 0.88; OpenAI reports parity with no statistically significant regressions on this set.
Both evaluation sets are described as deliberately difficult and not prevalence-weighted; the production set is built around cases where earlier Advanced Voice Models did not give ideal responses, and the results exclude system-level interventions such as Trusted Contact, localized crisis helplines and Parental Controls.
The GPT-Live-1 comparison values are lower than those printed in the July 2026 GPT-Live card (synthetic mental health 0.70 here versus 0.85 there; synthetic emotional reliance 0.86 versus 0.93; production mental health 0.78 versus 0.90); the new card states that comparison values come from the latest versions of previously launched models and that earlier published results are not directly comparable.
GPT-Live-1.1 is the default voice model for Plus and higher tiers (including Codex and Dots); GPT-Live-1.1 Quip is the default for other users.

Methodology Notes

Developer self-report. Production-prompt evaluations use real audio from users who opted to share voice interactions, passed through permission, deletion/opt-out and PII-reduction steps, transcribed, then graded. Synthetic evaluations generate policy-targeted text prompts, convert them to speech and use them as audio input. Scores are reported as decimal rates per category without item counts or confidence intervals; the significance statements are OpenAI's. Alignment evaluations were converted to audio on the hypothesis that GPT-Live inherits behaviour from its underlying 'thinking' models (GPT-5.6 Terra named as the comparison). The page says 'Published October 15, 2026'; the 8-page PDF's server Last-Modified is 2026-10-16 07:20 UTC. Curator re-verified 2026-10-19: page HTTP 200 ('Published October 15, 2026'); PDF HTTP 200 (305,977 bytes; Last-Modified 2026-10-16 07:20 UTC); Tables 1 and 2 re-read with the four-column header counted; July comparison values taken from the held GPT-Live card row.