What GPT-4 Is and What Changed From GPT-3.5
GPT-4 is OpenAI's fourth-generation large language model, released in March 2023. The jump from GPT-3.5 to GPT-4 was substantially larger than the jump from GPT-3 to GPT-3.5. GPT-4 added multimodal input (text and images), dramatically improved instruction-following, significantly better reasoning on complex multi-step problems, and reduced hallucination rates on factual queries. It also introduced a much larger context window. The practical effect: tasks that produced frustrating output from GPT-3.5 — detailed legal analysis, complex code reviews, nuanced multi-constraint writing — became genuinely workable with GPT-4.
Multimodal Input: What It Enables
GPT-4 Vision (GPT-4V) can process image inputs alongside text — you can send a photo, screenshot, diagram, or chart and ask questions about it. This enables use cases that were impossible before: reviewing a whiteboard diagram, analyzing a graph from a PDF, describing an image for accessibility purposes, or helping with visual debugging of UI issues. The image understanding is integrated into the model's reasoning rather than handled as a separate pipeline, which means it can reason about the combination of visual and textual content in a single response.
GPT-4's Performance on Professional Benchmarks
GPT-4's benchmark performance was genuinely remarkable at release. It scored around the 90th percentile on the Bar exam (GPT-3.5 scored around the 10th), performed similarly on the USMLE medical licensing exam, scored well on the SAT, GRE, and AP exams across multiple subjects. These aren't cherry-picked results — they represent broad domain knowledge and reasoning capability that GPT-3.5 lacked. The practical implication is that GPT-4 can engage meaningfully with domain-specific professional tasks — legal analysis, medical information synthesis, financial modeling explanation — where earlier models produced output too shallow to be useful.
GPT-4 Turbo and the Model Evolution
OpenAI released GPT-4 Turbo in late 2023, addressing the original GPT-4's two biggest limitations: slow speed and high cost. GPT-4 Turbo is faster and significantly cheaper while maintaining similar capability, making it practical for applications that weren't cost-effective with the original. It also extended the context window to 128K tokens. GPT-4o ('omni'), released in 2024, added native audio and image generation on top of vision, making it a true multimodal model rather than primarily a text model with vision added on.
Limitations: What GPT-4 Still Gets Wrong
GPT-4's improvements over GPT-3.5 are real and significant, but the model still has fundamental limitations. It hallucinations, though less frequently — factual claims in niche domains should still be verified. It has a knowledge cutoff date and can't access current information without web browsing plugins. It can still fail on multi-step math problems that require precise calculation rather than pattern-matching. And despite impressive benchmark scores, it can be confidently wrong in ways that require domain expertise to catch. Use it as a powerful thinking partner, not an infallible authority.