Lesson 5 of 8 · 8 min read · last verified 2026-08-26
Real work for nothing
In this lesson you will:
- Get substantial work done without paying
- Judge when a locally-run model is a genuine option
A course that is free forever should be straight about this: a great deal of useful AI work costs nothing, and the pressure to subscribe is not always matched by need.
What free tiers actually cover
Most major assistants have a free tier, and for a large share of ordinary use it is enough.
What you usually give up: a message cap, a smaller or older model once the cap is hit, queueing at busy times, and fewer extras — file uploads, image generation, longer context.
What you usually keep: the ability to do the everyday work in E3, E4 and E5. Drafting, tidying, explaining, summarising, rewriting.
If your use is a handful of tasks a day, test whether the free tier covers them before assuming it does not. E2·L9 gives the method for deciding between tiers; this lesson is about not skipping the free option in the first place.
Stretching a free tier
Everything in L2, with more effect:
Batch related work. One session, several tasks, rather than opening a new one each time.
Ask well the first time. Every retry spends a message. P1·L1’s examples pay for themselves fastest here.
Keep threads short. Caps that count messages punish long conversations directly.
Use more than one free tier. Several assistants have one. Using different providers for different tasks is legitimate, and it doubles as E2·L2’s comparison habit.
Running a model on your own machine
Genuinely possible now, and worth understanding even if you decide against it.
Several capable open-weight models can be downloaded and run on an ordinary computer through free desktop applications. Modern phones can run small ones. No account, no per-token charge, no message cap.
What it costs instead: several gigabytes of disk, initial setup effort, and a device with enough memory — a recent laptop with 16GB is a reasonable floor, and it runs faster with a decent graphics card.
What you get: capable models, noticeably below the largest hosted ones on hard reasoning, and completely adequate for summarising, drafting, tidying, extraction and classification.
The reason that is not money
Privacy is the strongest argument, and it is not really a cost argument at all.
A locally-run model does not send anything anywhere. That changes what you are permitted to put in — E4·L8’s paste test is answered differently when the text never leaves your device.
For anyone handling confidential material, client data, medical or legal information, or working under a policy that forbids external services, this is frequently the only compliant option. It also works with no internet at all.
If that is your situation, the setup effort is not a cost; it is the price of being able to do the task legitimately.
When local is the wrong answer
Even-handedly:
- You need the strongest available reasoning.
- You need very long context.
- Your device cannot run it, and buying hardware to save a subscription is false economy.
- You would spend more time maintaining it than the subscription would cost — L4’s arithmetic.
Try it now (6 minutes)
Check what your current free tier allows: how many messages, which model after that, what resets when.
Most people have never read it and are either paying for something they did not need, or hitting a cap they could plan around.
Check your understanding
Recap
Free tiers cover a great deal of ordinary work — check yours before assuming otherwise, and stretch it by batching, prompting well first time and keeping threads short. Running a model locally is a real option, costing disk, setup and some capability instead of money. Its strongest argument is privacy rather than price: nothing leaves the device, which changes what you are allowed to ask.
🗂 3 flashcards from this lesson join your daily review.
Previous: Time is the other budget · Next: Paying from where you are