GPT-4o

GPT-4o was the May 2024 release that established native multimodal as the closed-frontier standard. "o" stood for "omni": a single pretrained backbone handling text, audio, and vision tokens, rather than separate models stitched together with adapters. The real-time voice mode pushed conversational latency below 400ms, a threshold past which interactions feel synchronous rather than turn-based, and the demo that accompanied the launch became the reference point for what AI voice products were expected to feel like. Pricing dropped to $2.50 per million input tokens and $10 per million output, a roughly 12x reduction from the original GPT-4. The open-weights world spent the second half of 2024 chasing this combination: Llama 3.1, Qwen 2.5, and DeepSeek V3 all positioned themselves as cheaper or comparable on text, but native multimodal at GPT-4o's latency remained out of reach until well into 2025. The model also became the default backend for ChatGPT free users, which made its capabilities and quirks visible to a global user base in a way no prior frontier model had been.