SAN FRANCISCO — OpenAI unveiled GPT-5 on Wednesday, a model the company describes as its most capable to date, featuring simultaneous real-time processing of video, audio, and text inputs with what chief scientist Dr Priya Nambiar called "genuine cross-modal reasoning."
Unlike its predecessors, GPT-5 does not treat vision, language, and audio as separate pipelines. Instead, the model was trained end-to-end on a unified token stream that encodes all three modalities together, allowing it to track a conversation while simultaneously analysing a live video feed and responding with synthesised speech in milliseconds.
Benchmark Performance
OpenAI published results on thirty-two public benchmarks showing GPT-5 surpasses human expert performance on medical licensing examinations, bar exam simulations, and graduate-level mathematics. On the MMLU reasoning benchmark, the model scored 94.3%, compared to GPT-4's 86.4%.
Independent researchers at Stanford's Human-Centered AI Institute noted the results are impressive but cautioned that benchmark performance does not always translate to real-world reliability. "The model still hallucinates," said Dr James Watkins. "It does so less, and it is better at flagging uncertainty — but it is not solved."
Safety and Deployment
OpenAI said GPT-5 went through six months of red-teaming before release, including novel protocols for testing agentic behaviour — scenarios where the model takes sequences of actions in the world rather than just generating text. The company said it had implemented new "circuit breaker" guardrails that interrupt task execution if the model detects it is about to take an irreversible action without explicit user confirmation.