GPT-5 Architecture Leak Reveals Mixture-of-Experts with 1.8 Trillion Parameters

Documents circulating in AI research communities suggest that OpenAI's upcoming GPT-5 model employs a sophisticated sparse Mixture-of-Experts (MoE) architecture. According to these materials, the model contains approximately 1.8 trillion total parameters but activates only around 200 billion during any single forward pass.

This architectural choice follows the trend established by models like Mixtral and Google's Gemini, where computational efficiency is achieved by routing tokens to specialized expert networks. The sparse activation means that despite the enormous parameter count, inference costs remain manageable.

The leaked materials also suggest significant improvements in multimodal capabilities, with native support for video understanding and real-time audio processing. The model reportedly achieves near-human performance on several professional certification exams.

OpenAI has maintained silence on the reports, consistent with their practice of not commenting on pre-release information. Industry analysts expect an official announcement within the next quarter.