The question “Claude or ChatGPT” comes up in the first ten minutes of nearly every training session, and almost always in a form nobody can answer honestly: which one is better. The honest answer is “depends what you do”, and today we’ll try to show what exactly it depends on.
There will be no benchmarks here and no pricing table. Both go stale faster than you can finish reading. What there will be is a list of things Claude can’t do, because without it the whole comparison is an advert rather than advice.
What Claude won’t do and ChatGPT will
Let’s start from this side, because it’s the shorter and more concrete one.
Claude doesn’t generate images. No photos, no illustrations, no graphics for social media – and it isn’t a question of worse quality, it’s the absence of the feature on every plan. It will draw you a chart, a diagram or an interactive visualisation, building them as HTML, and Claude Design, a separate feature within Claude, will put together a prototype, slides or a one-pager. Useful at work, but no substitute for a picture of a cat in a hat.
Continuous voice conversation only speaks English properly. The kind where you just talk and listen, with no pauses and no key to hold, runs in claude.ai in English. Support for other languages is in beta, and you pick the language by hand rather than having it detected – so if you think in something other than English, this is where ChatGPT is in a different league. Dictation is a separate matter: in the terminal, /voice handles other languages and works well for dictating instructions, but that’s dictation, not conversation.
A free ChatGPT account gets you more of what isn’t chat. The message allowance is roughly comparable; the difference is in range: images and voice are there from the start, while Claude has no image generation at all and keeps Claude Design behind a paid plan.
There is also a difference in tone: ChatGPT’s answers tend to be looser. That one is a matter of taste, though, and we won’t pretend it can be settled.
The terminal isn’t the differentiator
Running in a terminal, on your own machine, with access to the folder you’re sitting in, gets presented as the thing that sets Claude Code apart. You don’t paste files into a chat window – you point at a directory and say what should happen inside it.
Except Codex from OpenAI does exactly the same. It runs in the terminal, reads and changes files, executes commands on your machine, and keeps its own instructions in AGENTS.md, the counterpart to CLAUDE.md. It can also split a job across several specialised agents. Both tools appeared at around the same time, so there’s no story here about one beating the other to it by years – the differences were always about how well each handles the job, not about whether it can reach your directory at all.
Where they do differ is what it costs to start. OpenAI includes Codex in every plan, the free one included – with the caveat that on Free and Go it does so temporarily, so check the current state before you build on it. Claude Code needs Pro or above, which means a paid subscription from day one.
So the question is how well each of them does the work.
Where Claude wins for us
Here we’re in the territory of what we observe in our own work rather than anything measured, and we want to say so plainly. There are three such areas.
The first is code quality on larger changes. Not “can it write a function”, which any of these models can, but whether the project still builds after a rewrite touching fifteen files, and whether it invented abstractions nobody asked for along the way.
The second is data analysis. With a spreadsheet running to a dozen tabs and formulas pointing at each other across them, what counts is patience in checking its own work, not how fast the first answer arrives.
The third, and the one that matters most to us, is working for hours on a single task. In a four-hour session on a proposal, a policy document or a database migration, what counts is whether the tool still holds to what you agreed in the first hour, and whether it will admit it doesn’t remember something instead of inventing a plausible-sounding sentence. That’s the difference between editing a text and catching your tool making things up.
An important caveat here: models on both sides change every few weeks and we won’t be surprised if this reads differently in six months. We’re describing what we see now.
To put it in one sentence: ChatGPT is an excellent assistant to keep in your pocket – ask it anything, have it produce an image, talk to it on the way to work. Claude comes into its own as the expert you sit down with for a big project and walk away from a few hours later with the thing finished.
When it’s simply what you’re used to
Honestly, at the end: in a large share of cases the choice between the two is habit and nothing more. If you’ve used one for a year, know how it answers and know how to phrase a request to get what you meant, that knowledge is worth more than the difference between the models. Switching means two weeks of worse results from a tool that isn’t objectively worse.
So the question is worth turning around. Instead of “which is better” – is there anything in your work you’re not doing because it can’t be done? If there is, and it’s about files, directories and repetitive operations on your own machine, that’s a real reason to try something else. If there isn’t, a leaderboard from the internet won’t change your day either way.