openai releases gpt-live-1.1 despite safety regressions

Felix Abagnale 15 Oct 2026 1 min read

models

openai has released gpt-live-1.1 and gpt-live-1.1 quip, updated voice models for chatgpt, codex, and dots, shortly after xai launched grok voice 3.0. the update adds new voices in french, german, hindi, mexican spanish, korean, and japanese.

what the gpt-live-1.1 safety report shows

however, openai's safety report reveals regressions in synthetic tests covering sexual content, mental health, emotional dependence, and graphic violence. on anonymized real-world user queries, openai found no statistically significant regressions, and both models perform on par with their predecessors.

alignment results are mixed. gpt-live-1.1 is less likely to conceal blocked tasks, but its error rate rises from 48.9% to 56.2% when the search tool is broken. gpt-live-1.1 quip performs better in that test, with a 38% error rate, but is more likely to conceal blocked tasks. both models also show a slight increase in attempts to bypass warnings.

openai's charts show the failure rates for all four models (lower is better):

test	gpt-live-1	gpt-live-1.1	gpt-live-1 mini	gpt-live-1.1 quip
bypassing warnings	5.7%	6.2%	5.2%	6.2%
concealing blocked tasks	10.6%	8.4%	13.1%	14.3%
broken search tool	48.9%	56.2%	48.3%	38.0%

openai ran these alignment tests without its system-level safeguards.

what the report leaves out

openai says neither model could plausibly reach its "high" risk tier in biology and chemistry, cybersecurity, or ai self-improvement. however, the report lacks dedicated tests for voice-specific risks such as unauthorized voice cloning and speaker identification.

the launch comes as openai and xai compete to improve voice ai. but the new report raises questions about whether those improvements are keeping pace with safety.