OpenAI released GPT-4 on March 14th. The announcement included multimodal capability, improved reasoning, and benchmark results showing human-level performance on professional exams. The model that ships is different from the preview in important ways.
The capability jump
GPT-4 is a qualitative improvement over GPT-3.5 on complex reasoning tasks. Where GPT-3.5 fails on multi-step problems that require tracking several constraints simultaneously, GPT-4 maintains coherent reasoning across longer chains. The improvement is most pronounced on: complex coding tasks, mathematical problem solving, nuanced instruction following, and tasks requiring integration of information from long context windows.
The multimodal announcement
GPT-4V (vision capability) was announced alongside GPT-4 text but not released publicly for months. The ability to provide images as input alongside text prompts is the announced capability. The API initially only accepted text. Vision capability became available to ChatGPT Plus subscribers in September 2023 and to API developers in October. The gap between announcement and availability was notable.
What actually changed for developers
The practical difference for API developers in March 2023 was: better code generation for complex tasks, more reliable instruction following, improved tool use, and better handling of edge cases and ambiguity in prompts. The 8K context window at launch was smaller than many expected. The cost was substantially higher than GPT-3.5 ($0.03 per 1K input tokens vs $0.002). Applications that benefited most from GPT-4's improvements were willing to pay the 15x cost premium.
The safety improvements
OpenAI spent six months on safety evaluations before GPT-4's release. The model is more resistant to jailbreaks and adversarial prompts than GPT-3.5, produces fewer toxic outputs, and is better calibrated to express uncertainty. These are real improvements but not absolute guarantees: sufficiently creative adversarial inputs can still produce outputs outside the guardrails. The safety improvement is a degree of difficulty increase for misuse, not a prevention.