订阅一个 AI 工具之前，我们核对的五件事（附 376 条记录的实测比例）

The short version first: turn "should I subscribe" into five checkable questions

The reason AI-tool subscriptions are hard to decide is not the number of options. It is that every vendor sells the same sentence: "start free, upgrade when you need it." That sentence contains nothing you can verify. While maintaining this directory (376 published records, each carrying the date it was last checked against the vendor's own site), we compressed the pay/no-pay decision into five questions. Every section below states both how we check it and how you can check it yourself, so you never have to trust our judgement.

1. Ask what the free tier is actually counted in

Of the 376 published records, 287 (76%) spell out a free tier in their price string, but only 211 (56%) state the tier together with a number; 303 (81%) carry a price string long enough (over 8 characters) to actually be readable. Across the whole directory that string averages 18 characters and tops out at 96 — treat it as a clue, not a ledger. The gap is entirely about the unit. Four units show up, and each fails differently:

• Per request: N runs per day. Fine for occasional lookups, cliff-like if you hit the cap.
• Per token: the default for model access. The same "free quota" burns at very different rates for Chinese and English text, so a token plan is not comparable across languages.
• Per minute/second: voice, video, meeting transcription. A 30-minute meeting costs ten times a 3-minute one; nobody tells you that on the pricing hero.
• Per seat: collaboration tools. A "free" plan with one seat means testing it with your team requires paying for the whole team.

How to check it yourself: look for "per month", "per year", "per seat" or "billed on usage" on the pricing page. If a plan says only "Free" with no unit, treat it as "no verifiable information" rather than "generous". Our rule in the directory is to copy the vendor's raw price string verbatim (for example Free / Plus $20/mo / Pro $200/mo) without beautifying or converting it, because converting it means smuggling our assumptions into your decision.

2. Convert the monthly fee into "cost per answer you kept"

monthly fee ÷ (kept answers per day × 30) — the denominator must be the ones you actually kept, not the number of generations. The difference is about an order of magnitude, since a large share of generations are exploration you discard.

Four raw price strings currently in the directory (all verified on 2026-08-29; the date is printed at the bottom of each record):

For token-metered tools, run the bill through /free-tools/llm-api-cost-calculator. It only multiplies the numbers you type; it never invents a business volume for you.

3. Check whether it is overpriced inside its own category

A price alone means nothing; a price band does. The 36 category pages (the taxonomy defines 41 categories) line up comparable tools. Sliced by primary category on 2026-09-18, over the union of dollar figures that appear in the price strings: writing and docs — 30 records, $5.75–$399, 25 of them quoting a price and 22 flagging a free tier; coding assistants — 25 records, $9–$100, 15 quoting and 17 with a free tier; video generation — 28 records, $5–$200, 25 quoting and 22 with a free tier. Having a free tier is not rare. What is rare is a free tier that survives one complete task, and reading "there is a free plan" as "it covers you" is the most common misstep in this decision. Two side-by-side pages make the same point in one screen: /compare/cursor-vs-github-copilot and /compare/chatgpt-vs-claude.

Three cases where paying above the band is defensible: you need private deployment or a data-retention commitment; you need batch or off-peak quotas; you need an SLA, invoices and procurement paperwork. Outside those, a $30+ premium that buys only a nicer interface is usually not worth it.

4. Read what it explicitly does not do

We require every record to carry both strengths and limitations: 376 of 376 have both. Separately, the data files hold 110 vendor-written "value proposition" strings that we deliberately do not render — sentences like "industry-leading intelligence" are marketing copy, and putting them on our page would mean endorsing them. That choice is worth more than 110 extra fields.

You can reproduce the check on any vendor site in three minutes:

1. Is there a public changelog? Products without one have the least durable feature claims.
2. Does pricing list an enterprise / security tier? If yes, the pricing page has survived real procurement.
3. Is there a human support entry point in a language you can read? Bot-only support means your incident latency is invisible on the pricing page.

5. Watch the verification date, not the promise

Price and free-tier text are the two fields that rot fastest. Our mechanism is dumb but auditable: every record has a lastVerifiedAt, and the current set falls between 2026-08-29 and 2026-09-15. Anything older than 180 days is named in the build log by npm run validate:tools (with a 30-day advance warning). That is not a claim of accuracy — it is a disclosure of how long it has been since a human checked.

Your own rule: re-open the pricing page within 72 hours of signing an annual contract and keep a screenshot. Annual-plan disputes almost always come from "the day I decided" and "the day I signed" being different days, and a screenshot is the cheapest insurance available.

What this checklist cannot do (stated plainly)

• What we copy is short by design: the price string averages 18 characters and tops out at 96. That is enough to decide whether to click through, not enough to compute an invoice — read the vendor page on the day you pay.
• We do not benchmark model quality. We publish no score of our own and we do not quote vendor scores as conclusions.
• We do not rank "the best". Ranking pages only answer "among the records here, which ones satisfy condition Y for scenario X" — the conditions are printed on the page so you can reject them.
• We cannot obtain discounts for you. Where a link is an affiliate link, the page says so; where we cannot say that honestly, we do not include the link.

One-sentence summary

Measure your own volume, read their unit, then divide the monthly fee by the answers you kept. If it still feels expensive after those three steps, it is expensive — and that has nothing to do with which model is smarter.