Key Features and Capabilities of OpenAI's GPT‑4

On April 1, OpenAI officially launched GPT‑4, a large multimodal model that accepts both image and text inputs and generates text outputs, achieving human‑level performance on many professional and academic benchmarks.

Multimodal capability: GPT‑4 can interpret images, enabling tasks such as describing picture content, analyzing charts, answering visual questions, and even generating code from hand‑drawn sketches, although image input is still under development.

Enhanced reasoning: The model shows markedly improved logical inference, exemplified by its ability to explain jokes involving outdated hardware and to solve complex problems with human‑like reasoning.

Longer context: GPT‑4 accepts up to 32,000 tokens (about 24,000 words) of input and can produce outputs of up to 25,000 words, opening possibilities for long‑form writing, extensive dialogues, and document analysis.

Exam performance: On standardized tests such as the GRE, SAT, and the bar exam, GPT‑4 scores in the top 10‑15% of human test‑takers, often matching or surpassing GPT‑3.5.

Multilingual support: Evaluated on the MMLU benchmark in 26 languages, GPT‑4 outperformed GPT‑3.5 and other large language models across all tested languages.

Integration with Microsoft Bing: Microsoft has already incorporated GPT‑4 into the new Bing preview, providing enhanced search experiences despite only text input being available.

How to access: Users can experience GPT‑4 by subscribing to ChatGPT Plus ($20/month), joining the waitlist for the GPT‑4 API, or contributing to the open‑source OpenAI Evals project.

Pricing and safety: The service costs $0.03 per 1k prompt tokens and $0.06 per 1k completion tokens, with rate limits of 40k tokens per minute; safety training reduces the likelihood of disallowed content responses to 18% and increases factual replies to 40%.

Conclusion: While OpenAI has not disclosed model size or training hardware, GPT‑4’s capabilities represent a significant leap in AI performance, prompting both industry competitors and researchers to reconsider the future of innovation.