I paid for 23 AI tools. Only 4 produced real work
I paid for 23 AI tools in 18 months. Only four produced real work. Here’s the audit, one-in-one-out rule, and 30-day test that cleared the clutter.
Eleven kinds of AI subscription. One question picks yours: coding, chat, task agents, always-on agents, images, video, music, voice, design and websites, research, and all-in-one bundles.
Answer a few questions and you get the plans that fit, ranked, with the monthly price and the reason each one made the cut. Every plan here is one you can actually sign up for with a normal card. Nothing leaves your browser: the catalog ships inside this page, so there's no account and no server watching your answers.
Catalog verified
Prices are the published dollar rates from that date. What a plan includes, and whether you can buy it at all, can differ by country.
Eleven tracks: coding, chat and writing, task agents, always-on agents, image, video, music, voice, design and websites, research, and all-in-one bundles. The first question picks your track, and the next three only ask what that track cares about. Nobody asking about song lyrics gets quizzed about tokens per day.
A task agent does what you ask, when you ask it. An always-on agent keeps a job: it has its own computer, holds context between sessions, and starts work without a prompt. Those are usually sold as a plan that already includes them (Google AI Pro, SuperGrok, a Cursor tier), not as a separate subscription, so that is how they are listed here.
API keys and pay-as-you-go credits bill by the pixel and never warn you. Fixed-price plans do. Region-locked deals you can't sign up for from outside that region are excluded rather than footnoted.
Volume fit counts most — and it is model-aware: your model answer is checked against each plan's per-model numbers, because that is how these plans actually meter. Then where you work, price efficiency, model fit, any tool you already pay for, and how many apps you connect. Half-answered forms get their score capped and labeled provisional. Answer everything and the cap comes off.
Because vendors meter by model, not by plan. Inside one $10 subscription, a flagship can be capped near a hundred requests per five hours while a workhorse open model effectively never binds — a several-hundred-fold difference at the same price. Simple mode uses your model-class answer (smartest / balanced / fast-and-cheap); Advanced mode lets you pick the exact model and uses the vendor's published per-model numbers where they exist, and measured community numbers where they don't.
Every record in the research catalog carries its own verification date and source links; the page shows each plan's vendor page and the catalog-wide verification date above (2026-10-11). Prices and quotas move fast, so treat the number as a starting point and confirm at checkout.
Not always. The figures are the vendors' published dollar prices on the day we checked. Local taxes, regional pricing, and which plans are offered vary by country, and we drop anything that only sells through a local wallet or phone number. Read the fine print at checkout.
Nowhere. The whole catalog ships inside this page and the matching happens in your browser. No account, no server call, no analytics event carrying your answers.
Picking between plans is half the fight. These go deeper on the other half.
I paid for 23 AI tools in 18 months. Only four produced real work. Here’s the audit, one-in-one-out rule, and 30-day test that cleared the clutter.
Too many productivity tools promise more output but create subscriptions, fake goals, and more work. Here’s the case for a smaller, honest tech stack.
Tech FOMO can turn an impressive tool into an imaginary requirement. Use this four-line filter to separate a real workflow need from more bloat.
I earn from affiliate links, but commission rates never decide my rankings. Here’s how to spot real testing, honest research, and payout-driven lists.