I pay for all three. Cursor Pro, Claude Pro, and ChatGPT Plus. Twenty dollars each, sixty a month total. I've run them side by side for months now, and the differences matter way more than the fact that they all cost the same.
Here's the honest breakdown: what each one is, what the limits actually are, and which one I'd keep if I had to cut the others.
They're not the same kind of tool
Before the limits, the shape of each one matters, because they solve different problems.
**Cursor** is an editor. It's basically VS Code with AI built in. You live inside it, and that's the point.
**Claude** is Claude Code, a terminal agent. It runs in your terminal, reads files, writes code, runs tests, and iterates in agentic loops.
**Codex** is OpenAI's coding agent, included in your ChatGPT Plus. It shows up in the CLI, the IDE, the web, and the ChatGPT desktop app.
So the real question isn't which is best. It's which fits the way you actually work.
The usage limits, which is what everyone asks about
Cursor Pro, $20
Cursor gives you two monthly usage pools. One for its own models (Grok 4.6, Grok 4.5, Composer 2.5). One for third-party models, billed at the provider's prices. Your included $20 is dollar-equivalent credits, so the expensive frontier models eat it fast.
The reset is monthly with your billing cycle, and unused usage does not roll over. Use it or lose it. When you run out, you either turn on on-demand pay-as-you-go or upgrade.
In practice, people report burning through roughly half a billion tokens, about $195 worth of usage, before it taps out. If you stick to autocomplete and Composer, it lasts. If you hammer premium models all day, it won't.
Claude Pro, $20
Claude Code is included on Pro. There are three ceilings, not one: a rolling five-hour session window, a weekly cap sitting above it, and a separate Opus-only weekly meter. None of them are message counts. It's token-based, and a long conversation re-sends its entire history on every turn, so a long thread drains quota faster than the message count implies.
Anthropic's rough public guidance is around 45 messages every five hours, but the window slides. Older usage ages out instead of waiting for a hard reset. When you hit a limit, Claude stops and names the reset time. Nothing is silently queued.
There's one escape hatch worth knowing: usage credits at standard API rates, with a spend cap you control.
Real-world, Pro sustains roughly one to two hours of active Claude Code a day. A heavy agentic loop can burn the whole five-hour window in 20 minutes.
ChatGPT Plus, $20
Codex usage is shared with ChatGPT Work. It's metered in estimated local messages per five-hour window, per model:
- GPT-6 Sol: 15 to 150 messages
- GPT-6 Astra: 5 to 45 messages
- GPT-5.6 Terra: 25 to 200 messages
- GPT-5.6 Luna: 250 to 2,000 messages
Weekly limits may also apply, and there's no published number for them. Cloud chats use Sol and can burn more of your allowance than local messages. At the limit, you can buy extra credits instead of upgrading.
One heads-up: GPT-5.5 retires from ChatGPT, Codex, and Work on October 14, 2026. If anything you built depends on it, migrate before then.
Advantages and disadvantages
Cursor
What it's good at: the best hands-on editing experience of the three. Autocomplete, inline edits, and tab completion that feels almost psychic. Fast for everyday work like refactors and small features. Its own model pool means Composer work doesn't burn your third-party budget.
Where it falls short: it gets expensive fast if you lean on frontier models. There's no rollover, so unused credits vanish. And out of the box it's less of a full autonomous agent than the other two.
Claude
What it's good at: Claude Code is the strongest agent here for real multi-file work. It reads a whole codebase and makes coherent changes better than the rest. Opus-class reasoning is available when a problem is genuinely hard.
Where it falls short: Pro's limits are the tightest of the three for heavy agent use, and the five-hour window can wall you mid-task. Claude chat and Claude Code share the same pool, so browsing claude.ai quietly eats your coding budget.
Codex
What it's good at: it's included in ChatGPT Plus, which a lot of people already pay for. It runs across CLI, IDE, web, and desktop with the same allowance everywhere, and the cheaper models are genuinely generous (Luna at 250 to 2,000 messages). You can buy credits instead of upgrading.
Where it falls short: there's no published weekly number, so you're partly guessing. Expensive frontier models drain the window fast, and it shares usage with ChatGPT Work.
So which one earns its keep
If you write code all day and want a single tool: Claude Code is the best agent, but Pro's limits will frustrate heavy users.
If you want the best editing experience: Cursor wins, but its credit model punishes frontier model use.
If you already pay for ChatGPT: Codex is the best value, and the cheap models give you the most messages per dollar.
What I actually do is keep all three for different jobs. Cursor for hands-on editing. Claude Code for agentic tasks that touch many files. Codex when I'm already working inside ChatGPT. Sixty a month sounds like a lot, and it is, but the alternative is picking one and hitting its wall on the exact day you need it most.
The twenty dollars is the same. The limits are not. Pick based on how you actually work, not which logo you like. And check the usage dashboard every week or two, because every one of these companies changes the numbers without announcing it.