OpenAI Unveils "GPT-4" with Multimodal Support

On March 14, 2023, OpenAI announced its new model “GPT-4.” Its standout feature was multimodal support — accepting images as input alongside text — and it marked a major performance leap over its predecessor GPT-3.5, scoring in the top 10% on a simulated bar exam. It became available to ChatGPT Plus subscribers the same day as the announcement.

Details

• Multimodal support: Accepts input combining text and images, enabling it to handle vision-and-language tasks together. Output remains text-only
• Performance gains: Scored in the top 10% on a simulated bar exam, a major improvement over GPT-3.5, which scored around the bottom 10%
• Rollout: ChatGPT Plus subscribers gained immediate access, subject to usage caps; API access opened progressively to developers via waitlist
• Safety work: OpenAI said it spent about six months tuning factuality, controllability, and guardrail adherence, drawing on internal adversarial testing programs and lessons learned from operating ChatGPT
• Knowledge cutoff: Trained on data up to September 2021, with an explicit caveat that it generally lacks awareness of events after that

What happened next

GPT-4 continued to receive updates afterward, including a Turbo version and expanded vision and voice capabilities, forming the foundation for GPT-4o in May 2024 and the reasoning-focused model o1 in September 2024. Many competing AI companies began using GPT-4 as a benchmark for performance comparisons, cementing its place as a turning point that dramatically raised the practical bar for generative AI.