What is ChatGPT?

What ChatGPT can do

ChatGPT is a general-purpose conversational AI assistant. Its core capabilities:

Writing and editing

• Draft emails, reports, proposals, articles, and presentations
• Edit for clarity, tone, grammar, and conciseness
• Summarise long documents or research papers
• Translate between languages (supports 50+ languages)

Coding

• Write code in Python, JavaScript, SQL, TypeScript, and dozens of other languages
• Explain what a piece of code does
• Debug errors and suggest fixes
• Refactor code for readability or performance

Analysis and reasoning

• Answer questions about documents you paste in
• Compare options and recommend based on criteria you specify
• Break down complex problems into steps
• Draft structured plans, outlines, and checklists

Creative tasks

• Write stories, scripts, poems, and presentations
• Brainstorm ideas and suggest variations
• Generate and edit images directly with ChatGPT Images (powered by GPT Image 2.5 at the time of writing), or write prompts for other image tools

The GPT model family

ChatGPT has been built on successively more capable models since launch, and OpenAI swaps the model underneath every few months. This is the lineup at the time of writing (September 2026); check the LLM Landscape 2026 for the live list.

How ChatGPT works

ChatGPT is built on OpenAI’s GPT models, which are transformer-based large language models. It generates responses by predicting the most likely next token given the full conversation history.

What makes it feel like a genuine assistant rather than a raw text predictor is the training pipeline:

1. Pre-training: The model is trained on hundreds of billions of words of text from the internet, books, and code. It learns language patterns, facts, and reasoning structures from this data.

2. Instruction fine-tuning: The model is further trained on human-written examples of question-answer pairs. It learns to follow instructions rather than just predict text.

3. RLHF and reinforcement learning: Human raters compare multiple responses and rank them. A reward model learns what responses humans prefer, and the LLM is tuned to produce higher-ranked responses. Current GPT models are also trained with reinforcement learning to “think” through a problem step by step before answering, which is what the reasoning-effort setting controls.

The result is a model that feels cooperative, helpful, and able to engage with almost any topic.

ChatGPT vs Claude vs Gemini

At the time of writing (September 2026). Model versions and prices change often; the LLM Landscape 2026 and ChatGPT vs Gemini vs Claude pages track the current state.

What ChatGPT cannot do

Recall recent events without web search: Every GPT model’s training data has a knowledge cutoff. Without web search enabled, it does not know what happened after that date.

Give reliable specific facts: ChatGPT can hallucinate. It can state statistics, URLs, names, and dates with confidence when the information is fabricated. Always verify specific factual claims from authoritative sources.

Take actions in the world without tools: The underlying model only generates text. ChatGPT can search the web, run code, and (on paid plans) use Agent Mode to operate a browser for multi-step tasks, but it cannot send your emails or update your company’s databases unless you connect those systems and give it permission.

Maintain memory across sessions by default: Each conversation starts fresh unless the Memory feature is switched on (available on Free, with more capacity on paid plans). Without it, ChatGPT does not remember that you told it your name last week.

Guarantee accuracy for high-stakes decisions: Legal, medical, and financial decisions require human expert review. ChatGPT’s output is a starting point, not a conclusion.

Using GPT models via the OpenAI API

Developers call the GPT models directly via the API to build custom products. This example uses gpt-6-sol, OpenAI’s mid-tier model at the time of writing; swap in gpt-6-luna for cheaper, faster answers:

from openai import OpenAI

client = OpenAI(api_key="YOUR_OPENAI_API_KEY")

response = client.chat.completions.create(
    model="gpt-6-sol",
    reasoning_effort="low",
    messages=[
        {
            "role": "system",
            "content": "You are an expert on Austrian tax law. Answer questions clearly and flag where the user should consult a registered tax advisor."
        },
        {
            "role": "user",
            "content": "Can I deduct home office costs if I work remotely three days per week in Austria?"
        }
    ],
    max_completion_tokens=2000,  # reasoning models count their "thinking" tokens here too
)

print(response.choices[0].message.content)

What’s next

• What is a Large Language Model? : The technology powering ChatGPT in depth
• Claude vs ChatGPT : Detailed comparison of Anthropic and OpenAI models
• ChatGPT Free vs Plus vs Pro : What each ChatGPT plan costs and unlocks
• What is AI Hallucination? : Why ChatGPT and other LLMs produce confident wrong answers

Further reading

• OpenAI Platform Documentation : API reference for developers using GPT models
• ChatGPT Enterprise : Business and data processing agreement details
• LLM Landscape 2026 : Full comparison of all major AI models
• What is Generative AI? : The broader category ChatGPT belongs to