The Free Tier Is Not a Contract: What to Do When AI Pricing Moves

A free AI tier is a marketing instrument, not a service-level agreement. Quotas, model access, and even which features stay free move on the vendor’s calendar. If a client deliverable depends on a zero-dollar allowance, a Tuesday changelog can stop the work. As of 23 August 2026, the pattern is not “every free plan vanished.” The pattern is that the shape of free access keeps changing, and small teams that treat the current shape as permanent get surprised.

Two recent moves show both directions. On 6 August 2026, OpenAI said Free and Go users would get unlimited everyday text chats on GPT-5.6 Luna, while file uploads, images, voice, and other tools stayed capped. In the other direction, Notion’s help pages now put full AI on Business and Enterprise, with only a trial allowance on Free and Plus. GitHub Copilot Free, checked 23 August 2026, is still a limited individual plan: official docs describe 2,000 completions a month and Auto model selection only. Cursor’s Hobby plan is still listed as free with limited Agent, Chat, and Tab use. None of these is a contract that your next month will look the same.

What is safe to bind to a free tier

Bind experiments, not the path that invoices a client.

Safe to park on free: trying a new model, drafting internal notes, one-off research, a weekend prototype, personal learning. If the vendor cuts the quota tonight, you lose time, not a deadline.

Do not bind to free: anything with a client date, a recurring production step, or data you cannot re-export in an afternoon. A weekly client report that only exists because Gemini’s free cap happened to hold, a booking bot that only answers because a free chat seat is up, meeting notes that assume Notion’s trial AI allowance will still run — those are operational dependencies dressed up as $0 line items.

The test is one sentence: “If this allowance disappeared on Monday, which paid job would slip?” If you can name a job, that tool is not free. It is an unbudgeted single point of failure.

A five-point dependency score

Score Core tools only (a client or weekly process dies without them). Trial tools can stay on free until they become Core. Use this starting rubric — adjust the cutoffs if your risk tolerance differs.

Each line is 0, 1, or 2:

1. Quota. 0 = published cap you stay under; 1 = published cap you already graze; 2 = unpublished, “fair use,” or you already hit it in a normal week.
2. Feature, not just seats. 0 = the job still runs on the free feature set; 1 = a nice-to-have moved behind pay; 2 = the capability you actually use (model choice, agents, export, commercial license) moved. Notion’s AI fold-in is this pattern.
3. Terms and region. 0 = you read the current terms this quarter and they fit the job; 1 = you skimmed them once; 2 = you have not read them, or the vendor can change region, training, or commercial use without notice.
4. Export. 0 = you have opened a JSON, Markdown, or CSV export; 1 = export exists but you have not tried it; 2 = the only copy is a screenshot.
5. Named substitute. 0 = paid plan or second vendor named with a price checked this month; 1 = you have a vendor name but no price; 2 = no name.

A starting rule of thumb, not a law: if a Core tool totals 8 or more, write a paid path or a documented fallback before the next billing cycle. A Trial tool can wait even at 10. Do not average the five lines into a false comfort score.

Three pieces of a real fallback

A fallback is not “we will figure it out.” It has a budget line, a migration path, and portable data.

Budget line. For each Core tool at 8+, write the cheapest paid plan that would keep the same job alive, checked on the official page this month. Starting examples, not quotes: ChatGPT Plus, Claude Pro, GitHub Copilot Pro, Cursor Pro. Put that number in the same sheet as rent and software. If the team cannot name the number, the free tier is already over-budget — you just have not been invoiced.

Migration path. One paragraph: where the work moves, how long a dry run takes, and who owns it. A workable starting dry run is: export recent prompts, replay a handful of real jobs on vendor B, compare error rate, switch the production shortcut. If you cannot run that dry run inside a normal work week, you do not have a path.

Data you can carry. Chats, prompt libraries, uploaded files, and any embeddings or automation graphs. Prefer tools that export JSON, Markdown, or CSV you have actually opened. A “download my data” button you have never clicked is not portability.

Free-tier movement also cuts both ways. ChatGPT’s August 2026 unlimited text chats are useful until a workflow needs images or file analysis — those still have separate limits. Treat each meter as its own product.

A 30-minute inventory

Set a timer. Do not optimize the sheet.

1. List every AI tool you opened in the last 30 days (10 minutes). Include “free inside another product”: Gemini in a Google account, Copilot in VS Code, Notion AI on Plus.
2. Mark each one Core or Trial (5 minutes). Core = a client or weekly internal process dies without it. Trial = you can skip it for two weeks.
3. Score the five dependency points on Core tools only (10 minutes). Open the live pricing or usage page. If you cannot open a number, leave the cell blank and treat that line as a 2 until you can.
4. Write one fallback line per Core tool (5 minutes): paid plan + price as of today, or second vendor + export format. If you cannot fill the line, that tool is the first one to replace or to budget.

Then price the substitute the way the cost-per-task guide would: monthly fee of the paid path, or (jobs × cost per job) if you would buy usage. A $20 seat that protects a $400 deadline is cheap. A $0 allowance that is the only way you deliver is not cheap; it is concentrated risk.

The free tier will keep moving. Your job is not to predict the next changelog. It is to make sure no paid week depends on a promise the vendor never signed.