OpenAI CEO Sam Altman revealed the latest progress on GPT-5 at a recent tech sharing event. According to internal sources, GPT-5's performance on complex reasoning tasks will improve several-fold over GPT-4, particularly in mathematical proofs, code generation, and scientific reasoning.
GPT-5 will reportedly introduce a novel reasoning architecture that enables deeper "thinking" processes before answering questions. This means the model won't merely generate answers based on pattern matching but will engage in multi-step logical deduction. In internal testing, GPT-5's MATH benchmark score jumped from GPT-4's 42% to over 85%.
Additionally, GPT-5's multimodal capabilities will receive a comprehensive upgrade, natively supporting video understanding, 3D scene parsing, and real-time audio interaction. OpenAI is also significantly improving the model's safety alignment mechanisms.
The release is expected between late Q3 and Q4 2024, with OpenAI first opening access to enterprise customers and API users before gradually rolling out to ChatGPT users.