Your $20 Plan Is Enough

Not a day goes by without me seeing a post on Twitter from someone asking for more usage out of their $100 or $200 plan, or for a reset. This topic has been on my mind for a long time, and I'd like to share why I think the $20 plan is the right amount for anyone working at a human pace, and why the top tier exists for something else.

Apart from the $500 plan ChatGPT recently added, which seems to be an exception (so far), companies generally offer three plans: $20, $100 and $200 (all USD). They all seem to describe them the same way: the $20 plan is "for everyday work", while the higher plans are "best for coding" or for "people who work with AI all day".

I've thought a lot about what those labels are supposed to mean. Either I'm using AI wrong, or the labels nudge people into thinking they need at least $100 to code or work with AI all day. Since companies survive by selling, I tend to believe it's the latter.

Either way, I can tell you that you can work with AI all day (whether that's coding, brainstorming, designing, etc.) on the $20 plan just fine, because that's exactly what I do. Full disclosure: I've tried all the plans that both ChatGPT and Claude offer, and I've settled on Claude's $20 plan.

I don't use the same model for everything. I pick the one that fits the task, all on medium effort:

• Haiku 5.5 for daily chats and online research.
• Sonnet 5.5 for most of the coding.
• Opus 5.5 for brainstorming, and for coding when my starting point is weak or unexplored.

Picking the right model for the job goes a long way. There's no need to use the biggest model to ask a quick question.

How I work

I start by brainstorming a feature or whatever task I need to do. I reflect and research to figure out which approach is best, then I start coding by setting up all the guardrails and rules the agent should follow. Coding is the last step, and it usually takes a while before I get there.

Reading the code matters just as much as brainstorming. Recently I asked Claude to wire up the front end and back end for a feature. Part of the back-end code it generated accepted an internal ID sent from the front end. That goes against every best practice, since anyone could change that ID and reach data that isn't theirs. If I had been vibe coding, it would have gone straight into the codebase. Being in the loop is what caught it, and it saved me from shipping a major bug and a security risk.

It doesn't matter whether it's a client's app or a personal one. You want the same security either way, and there's only one way to be sure: read the code.

I've noticed that people are reflecting less and less before acting, which leaves little room to catch a bug or a wrong setup that later requires more work. AI does most of the coding, yet I still write code myself when it can no longer follow my instructions.

This process involves me at every step and every change. AI models do the work in seconds, but then they wait, and they wait for me. I am the slowest part of the process, and I'm happy to be.

If you work this way, it's hard to use up the entire $100 plan, let alone the $200 one.

So who needs more?

To be fair, usage isn't just about how many messages you send. Anthropic says that file size, conversation length and model choice all affect how fast you reach your limit. If you regularly work with very large files or very long sessions, you may genuinely need more, even at a human pace.

But that's not who I see complaining. The people unhappy with the $200 plan are usually running agents nonstop, with little or no supervision. They let AI generate code from nothing more than an idea, and the only check they do is on the visual result. Do that in parallel, and the $200 plan will look like a joke to you.

That's what the top tier is really built for. It's a valid way to work, but it's a different way to work, not just "coding".

If you're the one reading, thinking and deciding, your $20 plan is enough.