Beyond text and images
OpenAI CEO Sam Altman announced GPT-5 in a livestream Wednesday morning, calling it "the first model that truly sees the world in motion." Unlike GPT-4, which processed static images, GPT-5 can handle live video input from webcams, drones, and smartphone cameras with latency low enough for real-time interaction.
In a demo, the model answered questions about objects moving across a room, tracked a person walking through a crowd, and identified a recipe being prepared on a kitchen counter. Altman said the model processes 30 frames per second and maintains context across several minutes of video.
Benchmarks and pricing
GPT-5 scores 92% on the new MMLU-Pro benchmark, up from GPT-4's 86%. On video-specific tests, it correctly identified actions in 94% of cases in the Something-Something v2 dataset.
Pricing starts at $0.05 per 1,000 tokens for text and $0.10 per minute of video processing. A premium tier with higher rate limits costs $200 per month. The basic free tier includes 10 minutes of video processing per day.
Safety and regulatory questions
The release raises new privacy concerns. OpenAI says all video data is encrypted in transit and deleted after processing unless users opt into training. But consumer advocates warn that real-time video analysis in public spaces could enable mass surveillance.
EU regulators said they will review GPT-5 under the AI Act before approving its deployment in member states. The UK's AI Safety Institute has already started testing the model's ability to recognize restricted objects and follow safety guidelines.