Most people still use ChatGPT the way they used the first version: type a question, skim the answer, close the tab. That is roughly a tenth of what the product does now. OpenAI has shipped memory, Projects, Canvas, reasoning models, Deep Research, voice conversations and a GPT store, and the interface barely changed to signal any of it.
I have been on the paid tier for months, using it daily for drafting, research, spreadsheet work and the occasional bit of Python I could not be bothered to write myself. Below is what genuinely earned a place in my workflow, what I quietly stopped touching, and where the whole thing still trips over its own feet.
The free tier covers more than most people assume
Before paying for anything, it is worth knowing what costs nothing. Free accounts get access to a capable general model, a small allowance of the reasoning model, limited image generation, file uploads and a handful of Deep Research runs. For someone answering emails or summarising a PDF twice a week, that is enough.
If you are weighing up free options across the board, this round-up of the best free AI tools judged by real user reviews is a useful sanity check before you hand over a card. The paid tier starts to matter when you hit rate limits mid-task, which happens faster than you would expect once you start leaning on it.
Custom instructions and memory: the quiet upgrade nobody talks about
These are the two settings that changed my output more than any headline feature. Custom instructions sit in the background and apply to every conversation. Mine are short and specific: I write for a UK audience, I prefer short paragraphs, I do not want bullet points unless I ask, and I want the model to flag uncertainty instead of hedging its way through an answer.
The effect was immediate. I was rewriting the same three things in every response before, and now I am not. If I had to put a number on it, maybe a third less editing on first drafts.
Memory is useful until it isn’t
Memory remembers context between chats: the client I am working with, the fact that I use a particular style guide, the project I mentioned last Tuesday. It is genuinely helpful for ongoing work. It also accumulates junk. After two months mine had decided I was “focused on launching a newsletter”, which was true for about a week and then quietly wrong for the rest. Open the memory settings occasionally and delete things. It takes two minutes and stops the model from confidently building on stale assumptions.
Projects turn the chat window into an actual workspace
A Project is a folder with its own instructions and files, and every chat inside it can see those files. I keep one per client: brand notes, the last four articles, a tone-of-voice document and a list of banned phrases. Opening a new chat inside it means no re-explaining, no pasting the same background information for the twentieth time.
For anyone juggling more than two ongoing pieces of work, this is the feature that justifies the subscription on its own.
Canvas is the one I resisted and now use most
Canvas opens a side panel where the draft lives, and you edit it there instead of scrolling through chat. The part that matters: highlight a paragraph, ask for a tighter version, and it rewrites just that paragraph rather than regenerating two thousand words and breaking everything else. You can ask for a shorter document, a longer one, a change in reading level, or inline comments explaining a fuzzy sentence.
For long-form writing this is a much better fit than the chat format. I stopped copying drafts into a separate editor for the first round of edits.
Voice mode is better than it has any right to be
Advanced voice lets you interrupt mid-sentence and it picks the thread back up. I use it on walks to think through article structures out loud, and to rehearse answers to interview questions. Latency is low enough that it feels like a conversation rather than a walkie-talkie. It is also startlingly good at detecting tone, so asking it to “be blunter” actually produces blunter responses.
Data analysis: the underrated workhorse
Upload a CSV and ask for cohort retention by signup month. I did this with a 6,000-row export, got a chart, a written summary and the Python code behind it. It flagged the two columns with missing values before I did.
The catch is arithmetic. It can chain calculations in ways that look plausible and are slightly off, particularly with messy date formats. Treat the output as a fast first pass and verify anything that ends up in a deck.
Reasoning models and Deep Research
Switch to a reasoning model for anything multi-step: debugging a formula, planning a project with dependencies, working through a decision with trade-offs. Answers take longer and are noticeably more careful. For a quick definition it is a waste of time.
Deep Research takes five to ten minutes and returns a long report with citations. I have used it for competitor scans and market overviews. It is a solid head start that saves maybe a day of skimming, though it does about 70% of the reading rather than all of it. Check the sources. Some are excellent and some are a blog post quoting another blog post.
GPTs and the store: mostly noise
Custom GPTs let you save a configured assistant with its own instructions and knowledge files. In practice, most of the store is a prompt wearing a logo. A handful are genuinely useful, particularly ones wired into a specific dataset. Building your own takes about fifteen minutes and beats browsing. I have one for editing that applies a fixed house style, and it is the only one I open regularly.
Where it still fumbles
It invents citations
Ask for a study on a specific claim and you may get a real-sounding author, journal and year that do not exist. This is less frequent than it was, but it has not gone away, and it is the single biggest reason to verify anything factual before publishing. The broader pattern across paid tools is covered in this accuracy review of paid AI tools, and ChatGPT sits mid-table rather than at the top.
Long documents drift
Give it a 60-page report and ask questions about page 47. It will usually find the answer and occasionally confuses two similar sections. Splitting the document into sections and asking targeted questions works far better than one giant prompt.
It agrees with you too readily
Push back on a correct answer and it will often fold. Ask it to steelman the opposite position explicitly if you want a real second opinion rather than a nod.
What the twenty dollars actually buys
Plus is $20 a month. You are paying for higher limits, the reasoning models, Deep Research, Projects, Canvas and priority access when the servers are busy. The breakdown of what each tier includes is worth reading before you commit, particularly if you are considering the $200 Pro plan, and this ChatGPT pricing review of the $20 tier is blunt about where the value sits.
Compared with the rest of the market, it is decent value. I tracked eleven paid AI subscriptions over six months and ChatGPT was one of three I renewed without hesitating, as you can see in this six-month performance review of eleven AI tools. If your work is mostly marketing copy at volume, a dedicated writing tool is the better buy, and the honest Jasper AI review explains why that niche still exists.
Who should pay, and who should wait
The subscription makes sense if two or more of these apply:
- You hit rate limits on the free tier more than once a week.
- You work on long documents where Canvas saves real editing time.
- You juggle multiple clients and want Projects to hold the context.
- You need Deep Research or reasoning models for analysis rather than chat.
- You use voice mode as a thinking tool rather than a novelty.
Hold off if your usage is occasional, if you mainly need image generation (the limits are tight), or if your employer already provides a workspace account. Students and light users can get a long way on the free tier plus one dedicated tool for the thing they do most.
The features worth learning properly are the boring ones: custom instructions, memory hygiene, Projects and Canvas. They do not make headlines, but they are the difference between a chatbot you occasionally open and a tool that quietly absorbs a chunk of your weekly admin. Start there, give it a fortnight, and the case for paying, or not paying, tends to make itself.