Gpt 5.4 starts reply with “Yes”

OpenAI's flagship model, GPT-4o, is displaying a peculiar and persistent bug affecting its Russian language output. Multiple users, including developers using the Codex integration in Microsoft's VS Code, report that the AI consistently begins its replies with the affirmation 'Yes' ('Да' in Russian). This occurs even when the user's prompt is an open-ended 'how' or 'why' question that does not solicit a yes/no answer, fundamentally breaking the conversational flow and the model's coherence from the first word.

The bug appears to be specific to the Russian language processing pipeline within GPT-4o's architecture, ruling out a simple memory or context-window issue in individual chats. Its presence in both the ChatGPT web interface and the VS Code extension suggests a deeper flaw in the model's response generation logic for that locale. While not a critical failure, this repetitive error degrades the user experience for Russian-speaking professionals and developers who rely on the model for drafting, coding assistance, and analysis, forcing them to edit or ignore the erroneous prefix in every response.

OpenAI has not yet issued a public statement on the cause or a timeline for a fix. The incident highlights the ongoing challenges of maintaining consistent, high-quality multilingual performance in large language models (LLMs), where subtle bugs can have a disproportionate impact on specific user groups. It serves as a reminder that even the most advanced AI systems can exhibit localized, illogical behaviors that require continuous monitoring and patching.