GPT-4

GPT-4 is the fourth model in OpenAI's GPT (Generative Pre-trained Transformer) series, a large multimodal model released on March 14, 2023 that accepts images and text as input and generates text. It passed a simulated bar exam with a score around the top 10% of test takers, where GPT-3.5 had scored around the bottom 10%, and it became the model behind ChatGPT Plus and Microsoft's Bing Chat. The GPT-4 Technical Report withheld the architecture, model size, hardware, training compute, and dataset construction, citing competition and safety, so GPT-4's parameter count has never been published.

Capabilities and Benchmarks#

GPT-4 scored 86.4% on the MMLU knowledge benchmark, against 70.0% for GPT-3.5, and when OpenAI translated MMLU into 26 languages, GPT-4 beat the English-language score of earlier models in 24 of them. The first release offered a context window of 8,192 tokens, with a 32,768-token variant, against 4,096 for GPT-3.5. OpenAI also reported predicting some of GPT-4's performance before training finished, using models trained with no more than 1/1,000th of its compute, an application of scaling laws.

The model still hallucinates. It scored 19 percentage points higher than GPT-3.5 on OpenAI's internal adversarial factuality evaluations, and its pretraining data ended in September 2021, so it lacked knowledge of later events. Its abilities were also uneven: a 2023 field experiment with 758 Boston Consulting Group consultants found that GPT-4 raised the quality of their work on tasks inside its capabilities and lowered accuracy on a task just outside them, an effect known as jagged intelligence.

Training and Alignment#

GPT-4 was pretrained to predict the next token on public and licensed data and then fine-tuned with reinforcement learning from human feedback (RLHF). OpenAI spent six months aligning the model before release and engaged more than 50 experts for adversarial testing, a form of AI red teaming. Alongside the usual reward model trained on human preference rankings, it added rule-based reward models: zero-shot GPT-4 classifiers that grade each response against a human-written rubric, rewarding refusals of harmful requests and answers to requests known to be safe.

Compared with GPT-3.5, the result responded to requests for disallowed content 82% less often and followed OpenAI's policies on sensitive requests, such as medical advice, 29% more often.

Image Input#

GPT-4 could read charts, screenshots, handwritten notes, and photos, and the launch demo turned a hand-drawn website sketch into working HTML. Image input was first offered through Be My Eyes, whose Be My AI feature describes surroundings for blind and low-vision users, and it arrived in ChatGPT in September 2023 as GPT-4V. GPT-4 with vision is a vision-language model (VLM), and open VLMs such as LLaVA used text-only GPT-4 to write the training conversations for visual instruction tuning.

GPT-4 answers in text and struggles with tasks that need precise spatial localization, such as drawing exact bounding boxes, so computer vision pipelines pair it with a dedicated object detection model. The Ultralytics LLM interface keeps both in one script: YOLO26 finds and localizes the objects in a frame, and the interface's image argument sends that frame with a prompt to an OpenAI-compatible vision model when the detections call for a written description.

Successors and Related Models#

GPT-3, released in 2020, was a 175-billion-parameter text-only model with a 2,048-token context. GPT-4 Turbo, announced in November 2023, raised the context to 128,000 tokens at a lower price. GPT-4o, released in May 2024, processed text, audio, and images in one network and answered speech in as little as 232 milliseconds, and GPT-4.1 brought a 1-million-token context to the API in April 2025. Reasoning models such as o1 followed a different path, trained with reinforcement learning to write a long chain of thought before answering, and GPT-5 combined fast and reasoning models behind a router in August 2025.

OpenAI replaced GPT-4 with GPT-4o in ChatGPT on April 30, 2025, and its deprecation schedule set October 23, 2026 as the API shutdown date for gpt-4 and gpt-4-turbo, with gpt-5.6-sol as the recommended replacement.