Last updated
ChatGPT is the page nobody needs to read to know what the product is, which is exactly why a description of it is worthless. The interesting thing about ChatGPT in its current form is that it is no longer really an assistant. It is a platform with an assistant on the front, and almost every serious question about it — is the subscription worth it, should we buy it for the company, do we need Claude as well — is a question about the platform rather than about how well it answers a prompt.
The chat box is the least differentiated part of the product. Every frontier assistant has one, they are all fluent, and the ranking between them changes often enough that choosing on conversational quality is choosing on noise. What surrounds the chat box is harder to replicate: file handling, a code-execution environment for data work, image generation, voice, connectors into other applications, custom assistants you can configure and share, and a mobile and desktop client that most competitors have not matched for polish.
That surface area is the actual product, and it is why the honest comparison against a rival is rarely "which one writes better." It is "which one already has the thing I need attached to it." For a large number of people the answer is ChatGPT by default, not because it is the strongest model on any given week but because the capability they wanted was already in the same window.
The single most underused thing in the paid product is uploading a document or a spreadsheet and asking questions of it. Analysis runs in a sandboxed environment where the model writes and executes code against your file, which means you can get charts, cleaned tables, and calculations from someone who has never opened a data notebook. For the many jobs where the data work is not hard but is annoying — reconciling two exports, finding the rows that do not match, summarising a quarter of survey responses — this closes a real gap.
It also fails in ways worth knowing before you rely on it. It is confident about what a column means, it will quietly drop rows it could not parse, and it does not know the business rules that make a number correct. Treat the output the way you would treat work from a fast, bright contractor who has never seen your data: worth having, worth checking. The same applies to long documents, where summarising is reliable and finding the one clause that matters is not.
A custom GPT is a saved configuration: instructions, some uploaded reference material, optionally a connection to an external service, wrapped in a shareable link. What you are buying is not intelligence but the elimination of setup — the analyst who re-explains the reporting format every Monday builds it once, and the rest of the team stops re-explaining it too.
The ceiling arrives quickly. A custom GPT is a prompt with attachments, not an application. It has no state between conversations, no reliable enforcement of the rules you wrote, and no way to guarantee that a user who asks sideways gets the behaviour you intended. Teams that expect an internal product from one tend to be disappointed; teams that expect a well-made shortcut are usually happy. If you find yourself wanting version control, tests, and audit logs around one, you have outgrown the format and want the API.
This trips up more people than it should. Paying for ChatGPT does not give you API access, and paying for the API does not give you the app. They are billed separately, metered differently, and aimed at different users: the subscription is a flat-rate consumer product with usage limits, the API is metered by tokens and has no interface at all.
The practical implication for anyone building something is that the app is where you work out whether an idea is viable and the API is where you ship it. Getting a prompt to behave in the chat window costs you a subscription you already have; running it ten thousand times a day is a line in an infrastructure budget. Our piece on token economics covers why that second number behaves so differently from the first, and how prompting is changing covers the first.
An individual asks whether the monthly fee improves their day. An organisation asks a completely different set of questions, and the answers are what actually decide the purchase: can we administer seats centrally, does it connect to our identity provider, can we control which connectors are enabled, is workspace data excluded from model training by default, what is retained and for how long, and can we produce an answer for an auditor about all of it.
The business and enterprise tiers exist to answer those, and the gap between them and an individual subscription is administrative rather than conversational — the model is not smarter, the controls are. The specifics of what each tier includes have been revised repeatedly, so confirm the current terms in OpenAI's own documentation rather than trusting any summary, including this one. What is stable enough to plan around is the shape: individual plans are consumer products with consumer data handling, and the business tiers are where the controls a security review asks about actually live.
The unglamorous part matters too. Shadow usage is the normal state of affairs in companies that have not bought anything — people paste work into a personal account because nobody gave them a sanctioned option. Buying a workspace is frequently less about enabling AI than about moving usage that already happens onto infrastructure you can see.
It is the wrong default when you need to verify rather than read. A general assistant produces fluent claims with no reliable provenance, and the browsing it does when prompted is not the same as a tool built around citation. If the output is going to be defended to someone, start with Perplexity instead — ChatGPT vs Perplexity separates the two cases.
It is the wrong default when the work is one very long document or one very large piece of code and the failure mode you fear is losing the thread halfway through. Plenty of people who use ChatGPT for everything else keep Claude open for exactly that; ChatGPT vs Claude and our longer write-up both land on the same split.
It is the wrong default inside a codebase. It will write good code for a problem you describe, and it has no idea what is in your repository. An editor-integrated tool like Cursor is not competing on model quality there; it is competing on knowing what the other four hundred files say.
And it is the wrong default if you only ever needed one narrow thing. A person who wanted transcription, or image generation, or grammar checking, and bought a general assistant to get it, has usually bought a worse version of a dedicated tool at a similar price. Breadth is only an advantage if you use the breadth.
Drafting, explaining, rewriting, summarising, working out what you actually think. This is the majority of usage and it is not exciting, but the fact that no tool switch is required is most of the reason it wins over marginally better single-purpose products.
A contract, a policy, a long report, a set of meeting notes. Summarising is dependable; locating the one clause that changes your decision is less so, which makes this an accelerant for reading rather than a substitute for it.
Upload the file and the model writes and runs code against it, returning cleaned tables and charts. The value is concentrated in the tedious middle of data work, and the risk is that it does not know which of your columns is authoritative.
Any request you re-explain weekly is a configuration you could have saved once and shared. The gain is removing setup, not adding capability, and it holds up well until somebody expects it to behave like an application.
Job descriptions, release notes, status updates, the third version of an announcement. Output quality matters less here than escape velocity from a blank page, which is the one thing a general assistant is unambiguously good at.
Developers use the app to find the phrasing that works, then move the stable version into code where it can be versioned and tested. Using the chat window as a workbench is cheaper than iterating against a metered endpoint.
ChatGPT has six tiers in 2026: Free ($0, with ads and tight limits in the US), Go ($8/mo, more volume but still ad-supported and missing advanced features), Plus ($20/mo, the sweet spot with full models and features), Pro ($200/mo, for power users wanting the largest context and highest Deep Research limits), Business ($25/user/mo), and Enterprise (custom). The pricing trap to watch: the free and Go tiers are noticeably degraded by ads and rate limits compared to a year ago, and the jump from $20 Plus to $200 Pro is steep with little in between — heavy users can outgrow Plus without Pro being worth 10x the cost.
For occasional questions, no. For anything where you would be annoyed to hit a limit mid-task, yes, and the reason is usually access to the fuller feature set rather than a better conversation — file analysis, image generation, and the higher-capability models are where the paid tiers separate themselves. The specific limits and what sits behind each tier change frequently, so check the current plan comparison rather than an article.
It depends on which product you are using, and this is worth getting right rather than guessing. Consumer plans and business or enterprise workspaces are handled differently, and the consumer settings include controls over whether your conversations can be used to improve models. Because these terms are revised periodically, read OpenAI's current documentation before making a policy decision for a team, and do not rely on a third-party description of it.
No. The subscription and the API are separate products with separate billing. The subscription gives you the applications and their usage limits; the API is metered by token usage and gives you no interface. If you are building something, expect to pay for both — one to develop against and one to run on.
Decide on the shape of your work rather than on which is smarter, because that ranking is unstable and the differences that matter are not. ChatGPT if you want the widest set of capabilities behind one interface and the best clients to reach them through. Claude if your work is long documents, long code, or prose you will publish with your name on it. Many people who are honest about their usage end up paying for both, and treat that as a reasonable cost rather than an indecisive one.
Trust it on shape and distrust it on specifics. It is reliable at structure, explanation, rephrasing, and telling you what kind of thing you are looking at. It is unreliable on figures, citations, dates, names, and anything it would have to have looked up. The failure mode is not vagueness but confident precision, which is the hardest kind of error to catch by reading.
Full review coming soon.