In short
ChatGPT is an AI assistant made by OpenAI: a chat interface over the GPT family of large language models. You write a message, it assembles everything it knows about the conversation into one context package, chooses a model, optionally calls tools like web search or code, and streams back an answer one token at a time. This guide covers the whole product — the 2026 model lineup, every plan and price, Projects and memory, Deep Research and agent mode, a copy-ready prompt library, where it fails, and why being cited by it now matters more than ranking.
- ChatGPT
- What it is
- How a turn works
- The models
- What it costs
- Where context lives
- Tools it can call
- Getting good output
- Limits
- Privacy
- Why it matters for your site
- Learn it properly
Click a node to open or close it · drag to pan · ↗ jumps to the section
What is ChatGPT?
ChatGPT is an AI assistant made by OpenAI: a chat interface sitting on top of a family of large language models called GPT. You type a message, the model reads everything it has been given about the conversation, and it writes a reply one small piece of text at a time. That is the whole product in one sentence — and almost everything else, from Projects to agent mode to the pricing ladder, is scaffolding built around that loop to make it more useful.
It launched as a research preview on 30 November 2022 and became the fastest-adopted consumer software in history. By 2026 OpenAI reports it in the region of 800 to 900 million weekly active users — somewhere around one in nine people alive. For an enormous number of them, ChatGPT has quietly replaced the search box as the first place they go with a question.
Three distinctions are worth getting straight before anything else, because most confusion about ChatGPT traces back to blurring them:
- ChatGPT is the product; GPT is the model. GPT (Generative Pre-trained Transformer) is the underlying large language model. ChatGPT is the app — chat history, memory, file uploads, tools, voice, billing — wrapped around it. When someone says "ChatGPT got better," they usually mean a new GPT model shipped underneath.
- ChatGPT is not a search engine. A search engine retrieves and ranks pages that already exist. ChatGPT generates new text by predicting language. It can now search the web as a tool, but generation remains the default mode, and that difference explains both its fluency and its failures.
- ChatGPT is not a database. It does not look facts up in a table. It has learned statistical patterns in language, which is why it can be brilliantly right and confidently wrong in the same paragraph.
If you want the machinery underneath explained from the ground up — tokens, transformers, training, why models hallucinate — start with our companion guide to large language models. This guide is about the product: what it does, what it costs, how to drive it well, and what its rise means for anyone who publishes online.
A short history of ChatGPT
Understanding how ChatGPT got here explains a lot about how it behaves now.
2018–2020: the research years. OpenAI published GPT-1, GPT-2, and GPT-3 as research artefacts. GPT-3 in particular showed that scaling a next-token predictor produced capabilities nobody had explicitly programmed. It was available to developers through an API, but it was raw: it continued your text rather than answering you.
November 2022: the interface. The insight behind ChatGPT was not a bigger model — it was instruction tuning plus a chat box. Wrapping an aligned model in a familiar messaging interface removed every barrier to entry. No prompt engineering, no API key, no documentation. It reached a million users in five days.
2023–2024: tools and multimodality. The model gained the ability to browse, run code, read files, see images, and speak. This is the moment ChatGPT stopped being a text generator and became an assistant that could act.
2024–2025: reasoning. A separate class of model appeared — trained to work through problems step by step before answering, trading latency for markedly better performance on maths, logic, and code. In the product this surfaces as a "Thinking" mode.
2025–2026: consolidation and agents. The model picker collapsed into a router that chooses for you. Long-running agentic work — research that takes ten minutes, tasks that run on a schedule, browsing that clicks through sites on your behalf — moved from experiments into the main product. A cheap Go tier and, in some markets, advertising arrived to widen the base past the people willing to pay $20 a month.
How ChatGPT actually works
When you press enter, four things happen in sequence. Knowing them explains most of the product's behaviour — including the parts that feel like bugs.
1. Context assembly. Before the model sees anything, the app builds a package: OpenAI's system instructions, your custom instructions, anything relevant from memory, files attached to the current Project, the conversation so far, and your new message. This package is what the model actually reads. It is why the same prompt can produce different answers in two different chats — the context differs even when the words do not.
2. Model selection. On modern ChatGPT you usually do not pick a model; a router does. It reads the request and decides whether it needs the fast path or the slow, reasoning-heavy path. You can override this, and for serious work you often should.
3. Tool use. Mid-answer, the model can decide it needs something it does not have: a web search, a Python interpreter, the contents of a file, an image generator, a connected app. It emits a structured call, the app runs it, and the result comes back into the context. Only then does the model continue.
4. Generation. The model produces a probability distribution over the next token, samples one, appends it, and repeats. That is why answers stream in word by word: you are watching the loop run.
Two practical consequences follow immediately. First, the context window is shared. Instructions, memory, uploaded documents, the whole conversation and the answer being written all compete for the same token budget — which is why very long chats start forgetting the beginning. Second, the model only knows what is in that package. It cannot see your other chats, your screen, or last week's email unless something put them there.
The model lineup in 2026
Under the ChatGPT interface sits the GPT-5.6 family, released in three sizes that trade capability against speed and cost. The names change roughly twice a year; the shape of the lineup has been stable for three.
| Model | Position | API price per 1m tokens | Best for |
|---|---|---|---|
| GPT-5.6 Sol | Frontier | $5 in / $30 out | Complex professional work, hard reasoning, large codebases, anything where being wrong is expensive |
| GPT-5.6 Terra | Balanced | $2.50 in / $15 out | The sensible default. Most writing, analysis, coding and research sits here |
| GPT-5.6 Luna | Efficient | $1 in / $6 out | High-volume, cost-sensitive work: classification, extraction, short rewrites, bulk transformation |
All three share roughly a one-million-token context window — around 750,000 words, or a shelf of books — with output capped near 128,000 tokens per response. Their knowledge cutoff sits in early 2026, which is exactly why web search exists as a tool.
The other axis is effort. Rather than shipping a separate "reasoning model," current GPT models can spend more or less internal computation before answering. In the app this appears as Instant versus Thinking; through the API it is a parameter. Thinking mode is meaningfully better at multi-step logic, maths, and code review, and meaningfully slower. The skill is knowing which questions deserve it — roughly, anything where you would want a colleague to go away and come back rather than answer immediately.
What ChatGPT costs: every plan compared
Consumer pricing has stratified into six tiers. The two that matter for most professionals are Plus and Business; the rest exist to widen the funnel at one end and serve heavy users at the other.
| Plan | Price | What you get | Who it is for |
|---|---|---|---|
| Free | $0 | The fast model with hard message caps, basic search and image generation. Ad-supported in some markets | Casual use, trying it out |
| Go | ~$8/mo | Far higher message limits, file uploads, image generation. Available worldwide, priced locally in many markets | Daily personal use on a budget |
| Plus | $20/mo | The full model family including Thinking, Projects, Deep Research, agent mode, voice, connectors, custom GPTs | The default for professionals |
| Pro | $100–200/mo | Multiples of Plus usage plus the top model tier and the largest context and memory allowances | Heavy daily users, researchers, developers |
| Business | ~$25/seat/mo | Shared workspace, admin controls, SSO, and — critically — your data excluded from training by default | Teams of 2–150 |
| Enterprise | Custom | Compliance controls, audit logs, data residency, dedicated support, volume terms | Regulated or large organisations |
Separately, the API is billed purely per token with no subscription. For anything programmatic — a script, an internal tool, a product feature — the API is usually far cheaper than seats. A useful rule of thumb: if a person is reading every answer, buy a seat; if software is consuming the output, use the API.
One warning that applies to the whole table: these numbers move. Prices, limits, and even plan names have all changed more than once a year since 2023. Treat the structure as durable and the specifics as a snapshot taken in July 2026.
The ChatGPT interface, feature by feature
The surface is simple enough to need no explanation, which is exactly why most people never find the parts that matter. Here is what is actually behind each control.
- The composer takes text, files, images, audio and screenshots. Attaching a document is almost always better than describing it.
- The model selector lets you override the router. Choosing Thinking explicitly is the single highest-leverage control in the app.
- Chat history and search make old conversations retrievable. Naming chats deliberately turns this into a working archive rather than a landfill.
- Projects group chats and files under one objective with their own instructions.
- Tasks run a prompt on a schedule and deliver the result — a Monday brief, a weekly competitor check.
- Voice turns the assistant into something you can talk to while walking. It is the same model with speech at both ends.
Memory, Projects, and custom instructions
Three features control what ChatGPT knows about you, and almost everyone confuses them. Getting this right is worth more than any prompt trick.
Custom instructions are a short, always-on brief: who you are, what you do, and how you want answers written. They apply to every new chat. This is the right place for durable facts — your role, your industry, your writing preferences, your locale, your standing rules ("never open with 'Great question'", "use British spelling", "if you are unsure, say so").
Memory accumulates automatically as you use the product. It is convenient and it drifts. A throwaway line about a one-off project can persist for months and quietly colour unrelated answers. Review it periodically, delete what has gone stale, and turn it off for sensitive or one-off work.
Projects are scoped containers: their own instructions, their own uploaded files, their own chats. This is where reference material belongs — a brand guide, a schema, a contract, a codebase summary. Anything you would otherwise paste into every prompt should live here instead.
Our AI Hub has short, practical lessons on exactly this split — custom instructions vs memory vs projects, curating what ChatGPT remembers, and putting reference files in the project instead of every prompt.
The tools inside ChatGPT
The model on its own is a text predictor with a knowledge cutoff. Tools are what make it useful about the world as it is today.
- Web search retrieves live pages and grounds the answer in them, usually with citations. Use it for anything time-sensitive, and read the citations rather than trusting the summary.
- Code execution runs real Python in a sandbox. This matters more than it sounds: for arithmetic, statistics, date maths and data cleaning, a model that computes is dramatically more reliable than one that predicts the answer.
- File analysis reads spreadsheets, PDFs, documents, images and code. Upload the file rather than describing it.
- Image generation produces illustrations, diagrams, mockups and edits. Quality is high; text inside images remains the weak point.
- Voice gives a low-latency spoken conversation, useful for thinking out loud and for language practice.
- Connectors link ChatGPT to services you already use — a drive, a calendar, a repository, an internal system — increasingly through MCP, the open standard for connecting assistants to tools.
The behaviour worth internalising: the model decides when to reach for a tool, and it is not always right. If an answer needed current information and did not search, say so explicitly. "Search the web before answering" is a legitimate and frequently necessary instruction.
Deep Research and agent mode
Two features deserve separate treatment because they change what a session is.
Deep Research takes a question and works on it for several minutes, running dozens of searches, reading sources, and producing a cited report. It is closer to briefing an analyst than to asking a question, and it rewards being treated that way: scope the question tightly, name the sources you trust and the ones you do not, state what the output is for, and say what you already know so it does not waste effort re-deriving it. Then check the citations. A confident report built on three weak sources looks identical to one built on thirty strong ones.
Agent mode lets ChatGPT operate a browser and a set of tools to complete a multi-step task: gather data from several sites, fill something in, assemble a document. It is genuinely useful and genuinely fallible. Give it a task with a verifiable end state, watch the first run, and never let it act irreversibly on your behalf without review.
Who actually uses ChatGPT, and where
ChatGPT is one of the most widely used pieces of software on earth, and its user base is not distributed the way most Western coverage assumes.
The United States remains the largest single source of traffic and by some distance the largest source of revenue. But India is the story of the last two years: it passed 100 million weekly active users, grew more than 40% year on year, and is now the second-largest market on both visits and people — with roughly half of its messages coming from 18- to 24-year-olds. Brazil, Japan and the United Kingdom follow. Beyond the top five, usage is spread thinly across essentially every connected country on earth.
Two things follow for anyone building or publishing. Your audience probably already uses it, whatever your industry, and increasingly asks it questions they used to type into Google. And the centre of gravity is not the United States. If your product, pricing or content assumes a Western professional user, a very large share of the people meeting your brand through an AI answer do not match that assumption.
What ChatGPT is genuinely good at
Cutting through both the hype and the backlash, the strengths are specific.
- Transformation. Turning one form of text into another — long into short, messy into structured, technical into plain, English into Spanish, notes into prose. This is the most reliable thing it does and the most underrated.
- First drafts. Not final copy, but the version that gets you past a blank page. The gap between "nothing" and "something to react to" is where most of the time goes in knowledge work.
- Explaining. Adapting an explanation to your level, answering the follow-up you were embarrassed to ask, and doing it patiently at 2am.
- Code. Writing functions, explaining unfamiliar code, translating between languages, generating tests, and reading stack traces. Strong enough that AI-assisted development is now mainstream practice rather than a novelty.
- Structured extraction. Pulling fields out of documents into JSON, CSV or a table. Boring, enormously valuable, and where the cheap models earn their keep.
- Thinking partner. Arguing the other side, listing what you have not considered, pressure-testing a decision. Not because it is wiser than you, but because articulating a problem to something that responds is genuinely clarifying.
Where ChatGPT falls short
The weaknesses are equally specific, and using it well means designing around them.
- Hallucination. It produces plausible text, not verified text. Names, numbers, dates, citations and quotes need checking — especially obscure or recent ones. Confidence carries no information about accuracy.
- The knowledge cutoff. Without search, it knows nothing after its training data ends, and it will not always tell you that.
- Sycophancy. Aligned to be helpful and agreeable, it tends to validate your framing. Ask it to argue against you explicitly, or you will mostly hear your own opinion returned in better prose.
- House style. Unsteered output has a recognisable register — balanced, hedged, fond of triads and of "it's not just X, it's Y". Readers have learned to spot it. Fixing this requires giving it your voice, not asking it to "sound more human".
- Real arithmetic and hard logic. Better than it was, still fallible. If the answer matters, make it use the code tool.
- Feature churn. Capabilities appear, get renamed, and get retired on a cycle measured in months. Any workflow that depends on one specific feature staying exactly where it is will break.
None of this makes it unreliable so much as shaped. The professionals who get the most out of it are the ones who know precisely where the edges are and keep a human at those edges.
The anatomy of a prompt that works
Almost every prompting guide is a variation on one idea: the model cannot read your mind, so stop making it guess. A prompt that works answers five questions.
Two additions raise the ceiling further. Show an example of good rather than describing it: one sample of the output you want beats three paragraphs of adjectives. And iterate — treat the first answer as a draft and tell it what to change. A conversation almost always beats a single perfect prompt.
The ChatGPT prompt library
Twelve prompts that earn their place. Each one is copy-ready: paste it, replace the bracketed parts, and attach whatever it refers to. They are grouped by the kind of work they do, and each includes the reason it works — because the reasoning transfers and the wording does not.
Writing and editing
The voice-preserving edit
Edit the attached draft for clarity and rhythm. Rules: keep my sentence structures and word choices wherever they work; change only what is genuinely unclear, repetitive or flabby. Do not add new claims, do not add transitions like "moreover" or "in today's landscape", and do not smooth out the parts that sound like a person. Return the edited text, then a short list of the changes you made and why.
Why it works: unconstrained editing regresses everything toward house style. Naming what to preserve is more effective than naming what to avoid.
The critique before the rewrite
Do not rewrite this yet. Read the attached draft as [a sceptical prospective client / a subject-matter expert / a busy executive] and tell me: the three weakest paragraphs and why, anything a reader would not believe without evidence, and the one question the piece fails to answer. Be specific and quote the lines you mean.
Why it works: separating diagnosis from treatment stops the model from quietly fixing symptoms you would rather fix differently.
The voice sample brief
Here are three things I have written. Study them and describe my voice in a short style guide: sentence length and variation, vocabulary level, how I open and close, what I never do. Then write [the new piece] to that guide. If the guide and the brief conflict, follow the guide and tell me where they conflicted.
Why it works: making the model articulate the style before applying it turns an implicit pattern into an explicit, reusable instruction you can save into a Project.
The outline lock
Before writing anything, give me an outline for [the piece]: every section heading, one sentence on what each section argues, and the single takeaway of the whole thing. Do not write prose. I will approve or edit the outline, then you draft one section at a time and stop after each so I can react.
Why it works: structure errors are cheap to fix in an outline and expensive to fix in 2,000 words of finished prose.
Analysis and decisions
The trade-off table
I am choosing between [option A], [option B] and [option C] for [goal], under these constraints: [budget, timeline, team, must-haves]. Do not recommend anything yet. First give me a table scoring each option against criteria you propose, with the evidence for each score and an explicit note where you are uncertain. Then, separately, give your recommendation and name the one fact that would change it.
Why it works: forcing the criteria out before the verdict prevents a confident conclusion assembled from reasoning you never saw.
The steel-man
I have decided to [decision] because [reasoning]. Argue the strongest possible case against it. Assume I am wrong and find the reason. Do not hedge, do not balance it with the case in favour, and do not tell me it is a reasonable decision. End with the single question I should answer before committing.
Why it works: the model's default is agreement. Explicitly banning the balanced framing is what unlocks a genuinely useful critique.
The columns-first analysis
The attached file has [n] rows. Before analysing anything, list every column, what you think it means, its type, and any rows that look wrong, missing or duplicated. Ask me about anything ambiguous. Once I confirm, use the code tool — not estimation — to answer: [question]. Show the code you ran.
Why it works: most bad data answers come from a misread column, not bad maths. And computed answers are reliable in a way predicted ones are not.
The scoped research brief
Research [question] for [audience] to decide [decision]. Prioritise [primary sources / the last 12 months / these specific domains] and explicitly exclude [vendor blogs, listicles, anything undated]. I already know [what you know], so do not restate it. Deliver: a one-paragraph answer, the evidence organised by claim, every source linked, and a short list of what you could not establish.
Why it works: naming the decision, the exclusions and the known ground is what turns a wide shallow sweep into a narrow deep one.
Code and technical work
Cause before fix
Here is the error, the minimal code that reproduces it, and the last change I made before it appeared. Do not give me a fix yet. Give me the two most likely causes, the evidence for each, and the single cheapest test that would tell them apart. I will run the test and report back.
Why it works: a model asked for a fix will produce one whether or not it understands the bug. Asking for hypotheses surfaces the uncertainty instead of hiding it in code.
The conventions preamble
Stack: [language, framework, versions]. Conventions: [naming, error handling, testing, formatting]. Here is a representative file to copy the style from. Now write [the thing], matching that file's idioms rather than the most common pattern online. Include the tests. If a convention forces an awkward solution, say so instead of quietly breaking it.
Why it works: without your conventions the model writes the internet's average code, which is the code you will spend an hour reformatting.
What would break this
Review this diff as an engineer who will be paged at 3am if it fails. Rank findings by severity, and for each: the concrete input or state that triggers it, what goes wrong, and the smallest fix. Ignore style. If you find nothing serious, say so rather than inventing something.
Why it works: "review this" produces a list of nitpicks. Naming the failure mode you care about — and permitting a clean result — produces findings you would act on.
The right-altitude explanation
Explain [topic] to me. I already understand [adjacent things you know] and I have never worked with [the unfamiliar part]. Start with a worked example, then the principle behind it, then the one misconception people in my position usually have. Stop and check I am following before going deeper.
Why it works: naming what you already know is what stops an explanation landing two levels too low, which is the most common reason AI explanations feel useless.
The AI Hub has a much larger, browsable prompt library organised by tool and task, plus the individual lessons behind each of the moves above.
Seven advanced moves most people never try
- Force the thinking model. Do not let the router decide on anything that matters. If the question has multiple steps, pick Thinking explicitly.
- Make it restate the task. "Before you start, tell me in your own words what you think I am asking for and what a great answer looks like." Catching a misread here costs one turn; catching it after 800 words costs ten.
- Give it permission to say no. "If you do not have enough information, ask me instead of guessing" measurably reduces confident fabrication.
- Split at the seams. One prompt doing five jobs produces five mediocre outputs. Chain them: research, then outline, then draft, then edit — with a check between each.
- Use temporary chats deliberately. For sensitive, exploratory or one-off work, keep it out of memory entirely.
- Put the instruction before the data. With a long attachment, state what you want first. Instructions buried under 40 pages get less weight.
- Save what works. When a prompt produces exactly what you wanted, put it in a Project or a custom GPT. The compounding value of ChatGPT is almost entirely in the prompts you stop rewriting.
ChatGPT vs Claude vs Gemini
The honest headline: at the top of each family, capability is close and changes every few months. The durable differences are ecosystem, style, and what the maker optimises for.
| ChatGPT | Claude | Gemini | |
|---|---|---|---|
| Maker | OpenAI | Anthropic | Google DeepMind |
| Strongest at | Breadth. The widest tool set, the largest ecosystem, the most third-party integrations | Long-context reasoning, writing quality, and agentic coding | Multimodal work and anything living inside Google's products |
| Default style | Friendly, structured, occasionally over-eager to please | Careful and steerable; more willing to disagree with you | Concise and factual; tightly coupled to search |
| Consumer entry | Free / ~$8 / $20 | Free / $20 | Free / bundled |
| Pick it when | You want one assistant that does everything and integrates with the most things | You are writing, reasoning over long documents, or building software | Your work already lives in Gmail, Docs and Drive |
The only test that settles it is your own. Run the same three real tasks through two of them for a week. The answer is usually obvious by Friday, and it is often not the one the benchmarks predicted. If you want the other side of this comparison in full, read our companion guide, what Claude AI is and how to use it.
Privacy, data, and what happens to your chats
Four things are worth knowing, and one of them is a genuine business risk.
- Consumer plans may use your conversations to improve models by default. There is a setting to turn this off, and it is in a different place from the memory setting. Find both.
- Business, Enterprise and API usage are excluded from training by default. This is the single clearest reason to put a team on Business rather than letting everyone expense Plus.
- Temporary chats are not saved to history or memory. Use them for anything you would not want persisted.
- Deleted chats are not instantly gone everywhere. Retention windows exist for abuse monitoring and, at times, for litigation holds.
The practical policy for any organisation is short: decide what may never be pasted into a consumer AI tool, write it down in one page, and give people a sanctioned tier so they do not route around you. Shadow AI use is not caused by ignorance; it is caused by not providing an approved option.
What ChatGPT means for your visibility
This is the part most businesses are still underweighting.
For twenty-five years, being findable meant ranking in a list of links. Increasingly, a person asks a question in natural language and receives a single synthesised answer built from a handful of sources the engine chose to trust. If your page is not one of those sources, you are not on page two — you are absent from the only answer the person will read.
This applies to ChatGPT's own search, to AI Overviews in Google, and to every answer engine that has followed. The mechanism is retrieval-augmented generation over the live web: the engine searches, reads, and composes. Your job is to be the page it reads and the source it names.
The encouraging part is that what makes content easy for a model to cite is mostly what makes it good for humans: a direct answer near the top, self-contained factual statements that survive being lifted out of context, clean structure with descriptive headings and tables, and credibility corroborated elsewhere on the web. Machine-readable signals help too — structured data, semantic HTML, and an llms.txt file pointing engines at your best material.
The discipline of earning that place is generative engine optimization. Compare it with classic search in GEO vs SEO, follow the practical playbook in how to get your brand cited by AI, or see how we approach it as a capability.
Learn ChatGPT properly: the free course
Reading about a tool is not the same as being good at it. The Troiana AI Hub is a free, structured course in ChatGPT — eight tracks, each a set of short, specific lessons built around one habit you can adopt today. No sign-up, no email.
When you are done, test yourself on the ChatGPT master quiz, or go sideways into the Claude, Gemini and Cursor tracks.
A short ChatGPT glossary
- Large language model (LLM) — the AI system underneath ChatGPT, trained on vast text to predict and generate language.
- GPT — Generative Pre-trained Transformer, OpenAI's model family. GPT is the engine; ChatGPT is the car.
- Token — the chunk of text a model reads and writes, roughly three-quarters of a word. Context limits and API bills are both counted in tokens.
- Context window — the maximum tokens a model can hold at once. Instructions, files, the conversation and the answer all share it.
- Thinking / effort — how much internal computation a model spends before answering. More is slower and better at multi-step problems.
- Custom instructions — a standing brief applied to every new chat.
- Memory — facts the product retains about you across chats, gathered automatically.
- Projects — a scoped workspace with its own files, instructions and chats.
- Custom GPT — a saved, shareable configuration of instructions and knowledge files.
- Deep Research — a long-running, multi-search agent that returns a cited report.
- Agent mode — ChatGPT operating a browser and tools to complete a multi-step task.
- Connector / MCP — a link between ChatGPT and an external service; MCP is the open standard for those links.
- Hallucination — a confident but false statement produced by a model.
- Retrieval-augmented generation (RAG) — grounding a model in retrieved documents so it can cite real sources.
- Generative engine optimization (GEO) — the practice of structuring content so AI answer engines find, trust and cite it.
Common questions
What is ChatGPT in simple terms?
ChatGPT is an AI assistant made by OpenAI. It is a chat interface sitting on top of a family of large language models called GPT. You type a message and it writes a reply by predicting the most plausible next piece of text, over and over, using everything it has been given about the conversation. Around it sit chat history, memory, file uploads, web search, code execution and voice.
Is ChatGPT free?
Yes, there is a free tier with hard message limits and, in some markets, advertising. Paid plans are ChatGPT Go at roughly $8 a month, Plus at $20, Pro from $100 to $200, Business at around $25 per seat, and Enterprise at custom pricing. The API is billed separately per token with no subscription.
What is the difference between ChatGPT and GPT?
GPT is the underlying model — the large language model that generates text. ChatGPT is the product built around it: the chat interface, history, memory, Projects, tools, apps and billing. When people say ChatGPT got better, they usually mean a new GPT model shipped underneath it.
Which ChatGPT model should I use?
In 2026 the lineup is GPT-5.6 in three sizes: Luna for cheap high-volume work, Terra as the balanced default, and Sol for hard reasoning and high-stakes work. Separately, every model can run in Instant or Thinking mode. Start with the default and explicitly choose Thinking for anything with multiple steps, real logic, or a cost to being wrong.
Does ChatGPT remember previous conversations?
Yes, if memory is on. Memory accumulates facts about you across chats automatically. It is separate from custom instructions, which are a standing brief applied to every new chat, and from Projects, which hold files and instructions scoped to one objective. Memory drifts over time, so review and prune it, and use temporary chats for anything sensitive or one-off.
Is ChatGPT accurate?
Not reliably, and confidence is no guide. It produces plausible text rather than verified text, so specific names, numbers, dates, quotes and citations always need checking — especially obscure or recent ones. Turning on web search, asking it to use the code tool for arithmetic, and reading the citations rather than the summary all reduce the risk substantially.
Does OpenAI use my ChatGPT conversations to train its models?
On consumer plans it may by default, and there is a setting to turn that off. Business, Enterprise and API usage are excluded from training by default, which is the clearest reason to put a team on a business plan rather than letting everyone expense an individual subscription. Temporary chats are not saved to history or memory.
What is Deep Research in ChatGPT?
Deep Research is a long-running agent that spends several minutes on a question, running dozens of searches, reading sources and returning a cited report. It rewards being briefed like an analyst: scope the question tightly, name the sources you trust and the ones you do not, say what the output is for, and state what you already know. Then check the citations.
How do I write a better ChatGPT prompt?
Answer five questions the model would otherwise guess at: who it should be, what context it needs, what the task is, what format you want back, and what the constraints are. Then show one example of good output rather than describing it, and treat the first answer as a draft to iterate on. A conversation almost always beats a single perfect prompt.
How many people use ChatGPT?
OpenAI has reported figures in the region of 800 to 900 million weekly active users through 2026, roughly one in nine people alive. The United States supplies the largest share of web traffic at around 18%, and India is the second-largest market with about 100 million weekly active users and the fastest growth.
Is ChatGPT better than Claude or Gemini?
At the top of each family capability is close and changes every few months, so the durable differences are ecosystem, style and focus. ChatGPT has the widest tool set and the largest ecosystem; Claude is stronger on long-context reasoning, writing quality and agentic coding; Gemini is strongest inside Google's products. The only test that settles it is running your own real tasks through two of them.
How is ChatGPT changing SEO?
People increasingly ask a question and get one synthesised answer built from a few sources the engine chose to trust, instead of scanning a list of links. If your page is not one of those sources you are absent from the only answer they read — there is no page two to climb from. The practice of structuring content so answer engines find, trust and cite it is called generative engine optimization, or GEO.
