What Happened
OpenAI released GPT-4, its most capable model at the time, as a multimodal large language model that could process both text and image inputs. GPT-4 was made available through ChatGPT Plus and the OpenAI API. Unlike previous releases, OpenAI disclosed almost no architectural details, citing competitive and safety concerns.
Why It Matters
GPT-4 represented a significant leap in reasoning capability. It passed the bar exam in the 90th percentile (vs. GPT-3.5's 10th percentile), scored in the 88th percentile on the LSAT, and achieved strong results across dozens of professional and academic benchmarks. Its multimodal capabilities — understanding charts, diagrams, and photographs — expanded the scope of what LLMs could do. GPT-4 became the backbone of ChatGPT Plus, Microsoft Copilot, and numerous enterprise AI applications.
Technical Details
- Architecture: Rumored to be a Mixture-of-Experts (MoE) model, though OpenAI did not confirm
- Parameters: Not disclosed
- Context window: 8,192 tokens (standard), 32,768 tokens (extended)
- Multimodal: Accepts text and image inputs, outputs text
- Training: Pre-trained on text and code data, then aligned with RLHF
- Benchmark results:
- Uniform Bar Exam: 90th percentile
- SAT Math: 700/800
- GRE Quantitative: 163/170 (80th percentile)
- Codeforces: Estimated 392 rating
- Safety: Extensive red-teaming and safety testing documented in a system card