GPT-4 Explained: Multimodal Input, GPT-4 Turbo, Plugins, GPTs and Why It Was Retired From ChatGPT (2026 Guide)

On 14 March 2023, less than four months after ChatGPT went viral, OpenAI announced GPT-4. If GPT-3.5 showed that people wanted to talk to AI, GPT-4 showed how capable that AI could become. It could read images as well as text, handle much harder reasoning, and it became the brain behind a wave of new ChatGPT features: plugins, browsing, Code Interpreter, image uploads and custom GPTs.

In this part of our "ChatGPT models explained" series, we explain GPT-4 in depth: what OpenAI said it was, what was new, what it could and couldn't do, how GPT-4 Turbo changed things, where it stands in October 2026, and what its story means for your career as a student or fresher in India.

What is GPT-4, and when did it launch?

OpenAI's announcement described GPT-4 as "a large multimodal model (accepting image and text inputs, emitting text outputs)." Let's unpack that:

• Large language model: like GPT-3.5, it is a transformer-based model pre-trained on a huge amount of text, then fine-tuned to be helpful.
• Multimodal input: "modality" means a type of data, such as text, images or audio. GPT-4 could take images and text as input. You could, in principle, show it a diagram and ask a question about it.
• Text output: it answered only in text. It did not generate images or speak by itself.

OpenAI also said GPT-4 was trained with reinforcement learning from human feedback (RLHF), the same family of techniques used for ChatGPT, and that the company spent six months making GPT-4 safer and more aligned before release, using lessons from adversarial testing and from ChatGPT.

The undisclosed size

A big change from earlier GPT papers: OpenAI did not reveal GPT-4's size. Its GPT-4 Technical Report said that, given "the competitive landscape and the safety implications of large-scale models", it would not share details such as the model's size, its architecture or the hardware used to train it. So any parameter count you see online for GPT-4 is speculation, not an official figure.

Interview tip: if someone asks "How many parameters does GPT-4 have?", the accurate answer is "OpenAI hasn't disclosed it." Knowing what is officially unknown is a sign of a careful engineer.

When it launched and who could use it

GPT-4's rollout happened in stages, which is typical for new frontier models:

• 14 March 2023: launch. GPT-4 became available to ChatGPT Plus subscribers (the paid plan launched on 1 February 2023), with a usage cap. Developers could join a waitlist for API access.
• Image input was not public at first. OpenAI said image inputs were "still a research preview and not publicly available."
• Context windows. The standard model handled 8,192 tokens, and a 32,768-token version (gpt-4-32k) had limited access. A token is a chunk of text, roughly part of a word, and the context window is how much text the model can consider at once.
• Free users stayed on GPT-3.5. For most of GPT-4's life, the free ChatGPT plan used GPT-3.5, so GPT-4 was one of the main reasons people paid for Plus.

What was new compared with GPT-3.5

OpenAI said that in casual conversation the difference between GPT-3.5 and GPT-4 could be subtle, but it showed up when a task became complex enough. GPT-4 was described as more reliable, more creative and able to handle much more nuanced instructions.

The headline example from OpenAI: GPT-4 passed a simulated bar exam with a score around the top 10% of test takers, while GPT-3.5 scored around the bottom 10%. OpenAI published results on many other professional and academic exams in its technical report too. Treat these as OpenAI's own tests, not proof that the model "understands" law like a lawyer.

In practice, students and developers noticed improvements in:

• Multi-step reasoning: maths word problems, logic puzzles and debugging.
• Following detailed instructions: formats, word limits, roles and constraints.
• Longer inputs: with bigger context windows, it could work with longer documents and code files.
• Other languages: OpenAI reported that on a translated version of a common knowledge benchmark, GPT-4 beat GPT-3.5's English score in 24 of the 26 languages it tested.

The GPT-4 era in ChatGPT: plugins, vision and GPTs

GPT-4 arrived alongside a burst of new ChatGPT features. Key dates from OpenAI's announcements:

• 23 March 2023: ChatGPT plugins. OpenAI introduced plugins that let ChatGPT connect to outside services, plus two plugins it hosted itself: a web browser and a code interpreter. Access started in alpha with a small number of developers and ChatGPT Plus users.
• July 2023: Code Interpreter (later renamed Advanced Data Analysis) became available to all Plus subscribers, letting ChatGPT run Python code on files you uploaded.
• 25 September 2023: "ChatGPT can now see, hear, and speak." Image understanding, powered by "multimodal GPT-3.5 and GPT-4", rolled out to Plus and Enterprise users, along with voice on mobile. This is when many people first used GPT-4's vision in practice, often called GPT-4V.
• 6 November 2023: OpenAI DevDay. OpenAI announced GPT-4 Turbo and GPTs, custom versions of ChatGPT that "combine instructions, extra knowledge, and any combination of skills." Anyone could build one without coding, and OpenAI announced plans for a GPT Store.

Plugins were later wound down in favour of GPTs, but the idea behind them, an AI model calling tools and external services, became central to today's AI agents.

GPT-4 Turbo explained

GPT-4 Turbo was announced at DevDay as a preview model for developers. According to OpenAI:

• It had a 128K token context window, which OpenAI said was equivalent to more than 300 pages of text.
• Its knowledge of world events ran up to April 2023, much later than the original GPT-4.
• It was cheaper to use through the API than the original GPT-4.
• A GPT-4 Turbo with vision version could accept images through the API.

A stable gpt-4-turbo model with vision followed in April 2024. Then, in May 2024, OpenAI launched GPT-4o, which became the main model in ChatGPT and gradually took over GPT-4's role.

Limitations: what GPT-4 still got wrong

OpenAI was clear that GPT-4 "still is not fully reliable." It said the model could "hallucinate" facts and make reasoning errors, and that people should be careful using its outputs in high-stakes situations. Other limits worth knowing:

• Knowledge cut-off. The original GPT-4 mostly lacked knowledge of events after its training data ended, so without browsing it could be outdated.
• Confidently wrong. OpenAI noted GPT-4 could be confidently wrong in its predictions and did not always double-check its work.
• Slower and limited. At launch, Plus users had message caps on GPT-4, and responses were slower than GPT-3.5.
• Text output only. It could read images but not generate them on its own.
• Bias and safety gaps. OpenAI's system card described remaining risks despite safety work.

The same rule applies to every model: use AI for speed, then check the facts yourself.

Current status of GPT-4 (October 2026)

• In ChatGPT: GPT-4 was retired from ChatGPT on 30 April 2025 and fully replaced by GPT-4o, which had become ChatGPT's default model. GPT-4o itself has since been retired from ChatGPT too (13 February 2026). Today's ChatGPT runs on GPT-6, rolled out on 7 and 8 October 2026, with Sol for paid plans and Luna for Free and Go users.
• In the API: GPT-4 stayed available to developers after leaving ChatGPT, but OpenAI's deprecations page now lists gpt-4, gpt-4-0613, gpt-4-turbo and gpt-4-1106-preview (plus fine-tuned GPT-4 models) with a shutdown date of 23 October 2026, announced on 22 April 2026. Older versions such as gpt-4-32k, gpt-4-vision-preview and gpt-4-0314 were already shut down earlier. If a tutorial tells you to call gpt-4, use a current model from OpenAI's documentation instead.

Career lessons from the GPT-4 story

GPT-4's story is really about AI becoming a platform, not just a chatbot. That points to specific skills:

• Multimodal thinking. GPT-4 showed that one model can work with text and images. Projects that combine both, for example reading a chart and explaining it, are strong portfolio pieces.
• Tool use and integrations. Plugins, Code Interpreter and GPTs were early versions of AI calling tools. Learning APIs, JSON and function calling prepares you for roles in AI engineering and backend development.
• Context management. Bigger context windows created new work: retrieval-augmented generation (RAG), chunking documents and choosing what the model should see.
• Evaluation and safety. OpenAI's six months of safety work and red teaming highlight roles like AI evaluation, trust and safety, and AI quality analysis.
• Reading primary sources. The GPT-4 Technical Report and system card are public. Reading them, and noticing what is not disclosed, is a habit that sets you apart in interviews.

Project ideas you can build with any current model:

• A "notes to quiz" tool that reads a photo of handwritten notes and creates practice questions.
• A mini custom assistant for your college club, with fixed instructions and a small knowledge file, like a simple GPT.
• A data analysis notebook where AI helps you clean a public Indian dataset, and you verify every step.

Describe these with clear outcomes on your resume, then check them with the Ucanly AI resume review.

Where GPT-4 fits in the ChatGPT family

GPT-4 is the second chapter of ChatGPT. It followed GPT-3.5, which launched ChatGPT in November 2022, and powered the paid ChatGPT experience through 2023 and early 2024. GPT-4 Turbo extended it, and GPT-4o (May 2024) replaced it as ChatGPT's main model, followed later by reasoning models, GPT-5 and today's GPT-6.

Read the full series on the Ucanly blog, and if you are new to AI tools, start with our guide on how to use ChatGPT and Claude.

What you should do next

1. Read OpenAI's GPT-4 announcement (14 March 2023) and skim the GPT-4 Technical Report. Note what it says, and what it chooses not to say.
2. Write a one-page timeline from GPT-3.5 to GPT-4 to GPT-4 Turbo to GPT-4o, with dates. It is great interview preparation.
3. Try a multimodal task in today's ChatGPT: upload a diagram from your syllabus and ask for an explanation, then verify it against your textbook.
4. Learn the basics of tool calling with any current AI API and build one small tool-using project.
5. Practise checking AI answers: pick ten tough questions, ask an AI, and grade each answer honestly.
6. Follow a structured path in AI, data or software development through our courses, and explore roles that use these skills on the jobs board.

GPT-4 has left ChatGPT and is leaving the API, but the shift it started, from a chat window to a platform that sees, uses tools and can be customised, is the world you are building your career in.