The AI market has the energy of a gold rush where the surest money is in selling subscriptions. I pay for a handful of tools and have cancelled more than I keep — which is roughly the correct ratio. Here is the framework I actually use, plus the specific categories worth real money and the ones almost never are.

The one test that decides everything

Before paying for any AI tool I ask: does the paid tier remove friction from work I already do weekly, or does it just promise a feature I'll try once? The free tier is your trial. If after two weeks of genuine use you have not hit the ceiling of the free version, you don't need the paid one yet.

Worth paying for: a tool that saves you an hour you would otherwise spend every single week. Everything else is a demo with a bill.

What earns its fee

• The writing/reasoning model you use daily: the premium tier of one mainstream chat model pays for itself if you use it for real drafting, summarizing long documents, or code — the time saved is concrete.
• Deep-research / browsing modes: valuable when you genuinely need sourced, long-form answers, useless as an impulse. I keep it only during projects that need it.
• Image + transcription tools with flat pricing: one-off credits you consume predictably beat "unlimited" plans that throttle you exactly when it matters.

What I stopped paying for

• Thin "GPT wrapper" apps: a login screen bolted onto the same underlying model as the free version. If its whole value is a slightly nicer interface, that is not a 20-a-month product.
• Multiple overlapping chat subscriptions: two frontier models is plenty; the third rarely gets opened after week one.
• AI "all-in-one productivity" bundles: the AI feature is the marketing hook; the real product is a project tracker you can get cheaper alone.

The trap of "unlimited"

Unlimited plans are real but come with rate limits that bite exactly when you lean on them. For a normal working person, a mid-tier metered plan or buying credits is cheaper and less frustrating than paying for a theoretical infinity you'll be throttled out of anyway. Read the fair-use section — the number that matters is hidden there.

How I keep the stack honest

1. Annualize mentally, then pay monthly first. Commit to a year only after a tool survives a full quarter of weekly use.
2. One "serious" model subscription. Everything else is free tier until it proves the paid tier.
3. A calendar reminder before each renewal. More money is wasted on forgotten auto-renews than on any price I actually compared.

The bottom line

AI tools are worth paying for when they plug into a loop you already run every week. The moment a subscription is selling you a possibility instead of saving you a recurring hour, it is a candidate to cancel. Keep one or two that clear that bar and let the rest stay free.

Honest disclaimer: model capabilities and pricing shift constantly — a plan that is not worth it today may be next quarter, and vice versa. This is a decision framework, not a recommendation to buy any specific product, and I have no affiliate interest in the tools I keep or cancel.