When free AI slows down, do the math first

A friend asked me last week if AI was broken. He opened Doubao 2.2 to write a weekly report, and the second half started lagging. I told him to check Usage in Settings; his free quota was nearly exhausted. The model didn't suddenly get dumber; the free lunch ended, and rate limiting kicked in.

As a reviewer, I habitually run things myself. Recently, I've tested Doubao 2.2, another newer model, and OpenAI on various tasks. My conclusion is: don't rush to subscribe; do the math first.

Preparation is simple. Open a memo or spreadsheet and create six columns: Date, Tool, Task, Quota Consumption, Experience, Must Pay?. Choose three types of tasks: writing weekly reports, organizing meeting minutes, and generating PPT outlines. Two or three tools are enough.

Run the free tier first. Enter a fixed prompt in the chat box: "Help me organize the following content into a three-paragraph weekly report...". When you see slowdowns or insufficient quota, don't close it yet; note the time.

Then go to the usage page to check consumption. Usually found in Settings, Billing, or Profile. You'll see remaining requests, Tokens, or request counts. Tokens can be roughly viewed as finer-grained character counts.

Next, compare different models. If the interface has a Model dropdown or "Light/Standard/Premium" options, feed the same question into each and observe completion quality, speed, and hallucinations. In my tests, light models are sufficient for weekly reports, but fall short for long-document summarization.

Finally, fill in the conclusion table. Only if a task repeats three times in a week AND the free tier frequently fails does it qualify as a necessity. Occasionally generating titles isn't worth paying for.

Reports indicate that many users encounter low-performance modes after exhausting free quotas; data also shows that while there are many AI users, only 9.8% actively pay.

More tools don't mean you should pay; more users don't mean YOU are willing to open your wallet.

There are pitfalls. The most common is calculating only chat usage, ignoring uploads. Casually dropping a PDF or image can consume much more quota than a text paragraph, quietly draining your balance. The second pitfall is defaulting to the strongest model; most weekly reports don't need it. The third is setting no limits when building small tools with APIs. Think of APIs as interfaces for programs to talk to each other; every request costs money. The fourth is carelessly pasting company data; anonymize whenever possible.

The conclusion is direct. This ledger turns "Did AI start charging?" into "Where exactly do I need it?" The downside is spending a few days recording. High-frequency fixed tasks justify payment or switching to lighter solutions; occasional copywriting can stay on the free tier.

Next steps: lock down three essential tasks and create a set of prompt templates. Further down the line, see if you can run a small agent in a Docker sandbox. Docker is like putting a program in a temporary room; an agent is an AI that can execute a series of small tasks autonomously. Finish the math, then decide whether to subscribe.