Article updated in November 2026.
Update, November 2026: this article was first written in May 2023, when GPT-4 was new and many details were still rumours. Several of the original predictions turned out to be wrong, so the text below has been corrected. A section near the end explains what has changed since and which models to use now.
What is GPT-4?
GPT-4 is a large language model from OpenAI. It was announced in March 2023 as the successor to GPT-3 and GPT-3.5, the models behind the first version of ChatGPT. GPT-3 had 175 billion parameters and was already good at generating text, but it had known limits. It made factual mistakes, handled long documents poorly, and was hard to steer. This article looks at what GPT-4 changed compared with the third version.
Model size: what we actually know
Before the release, many sources said GPT-4 would have 10 trillion parameters or more. That figure was a rumour. OpenAI has never published the parameter count or the architecture of GPT-4. The company said only that the model is larger and more capable than its predecessors. Any specific number you see for GPT-4 is an estimate, not an official fact.
The lesson is that parameter count is not the best measure of quality. Training methods, data quality, fine-tuning with human feedback and safety work all affect how useful a model is.
Training data
The original article said GPT-4 would be trained on Wikipedia, social networks and other online platforms. OpenAI has not published a detailed list of sources. It says the model was trained on a mix of publicly available data and data licensed from third parties, then refined with reinforcement learning from human feedback (RLHF). So the claim of a specific set of sources was not confirmed.
Multimodality: images in, text out
This is the biggest real difference from GPT-3. GPT-4 accepts images as well as text as input. It can describe a photo, read a chart, explain a meme or work through a diagram. Its output is still text.
The 2023 prediction that GPT-4 would generate photorealistic images and video was incorrect. GPT-4 itself did not do this. Image and video generation came from separate OpenAI products, such as DALL·E and later Sora, and more recently from image generation built into newer ChatGPT models.
Better control over the output
GPT-4 follows instructions much more reliably than GPT-3. In practice you can set:
- the topic and the target audience;
- the style and tone, for example formal, friendly or technical;
- the length and structure, such as a list, a table or a fixed number of paragraphs;
- a role for the model through a system message in the API.
This steerability is a large part of why GPT-4 became useful for business tasks and not only for experiments.
Other real improvements over GPT-3
- Reasoning and accuracy. GPT-4 scored far higher than GPT-3.5 on many professional and academic exams. It still makes mistakes and can state false information confidently, so important facts need checking.
- Longer context. At launch GPT-4 worked with context windows of 8,000 and 32,000 tokens, much more than GPT-3.5. Later versions went much further.
- Safety. OpenAI reported that GPT-4 was less likely to answer requests for disallowed content than the previous models.
- Languages. Quality in languages other than English improved noticeably.
Speed
The original article predicted that GPT-4 would be faster than GPT-3. At launch it was the opposite. GPT-4 was slower and more expensive to run than GPT-3.5, which is why the faster GPT-3.5 stayed available for simple tasks. Speed and cost improved later with newer models such as GPT-4 Turbo and GPT-4o.

Applications
GPT-4 is used for many kinds of text work:
- copywriting and drafts of articles;
- social media content and news summaries;
- customer support assistants and chatbots;
- help with programming, from explaining code to writing and fixing it;
- analysing documents, tables and images.
Output should be reviewed by a person, especially for facts, numbers and legal or medical topics.
GPT-3 and GPT-4 compared
| Feature | GPT-3 / GPT-3.5 | GPT-4 (2023) |
|---|---|---|
| Input | Text only | Text and images |
| Parameters | 175 billion (GPT-3) | Not disclosed by OpenAI |
| Context window | Several thousand tokens | 8,000 and 32,000 tokens at launch |
| Instruction following | Limited | Much better |
| Speed | Faster | Slower at launch |
What has changed since 2023 and what to use now
GPT-4 is no longer the current model. In short:
- In late 2023 OpenAI released GPT-4 Turbo, with a much longer context window.
- In 2024 came GPT-4o, a faster model that handles text, images and audio, and then the o-series reasoning models.
- In 2025 OpenAI introduced GPT-4.1 and then the GPT-5 family. GPT-5.1 and GPT-5.2 followed, and GPT-5.2 became the default model in ChatGPT.
- OpenAI retired the original GPT-4 from ChatGPT in 2025. In February 2026 it also retired GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini from ChatGPT.
If you use ChatGPT, you now get the latest default model automatically. If you build on the API, check OpenAI’s documentation for current model names, prices and deprecation dates. Older models are removed on a schedule, and the best model for your task changes often. GPT-4 is still useful as a reference point for how modern language models developed.
Useful links
Here are some useful links for those who want to learn more about GPT-4:
- OpenAI’s announcement of GPT-4: https://openai.com/blog/announcing-gpt-4/
- VentureBeat article on possible applications of GPT-4: https://venturebeat.com/2023/05/01/how-openais-gpt-4-could-transform-content-creation-and-more/
- Official OpenAI website: https://openai.com/
Conclusion
GPT-4 was a real step forward from GPT-3. It followed instructions better, reasoned better, accepted images and handled longer texts. Several pre-launch expectations were wrong: the 10-trillion-parameter figure was never confirmed, GPT-4 did not generate images or video, and it was slower than GPT-3.5 at launch. Newer models have replaced it, so use the current ChatGPT default or the latest API models for new work.
Leave a Reply