Hope Is Not a Receipt

2. Multipliers without a measure

Several providers sell multipliers. OpenAI sells Pro tiers described as 5x and 20x the usage of Plus, while Claude Max offers 5x and 20x the session capacity of Pro. Cursor also markets agent limits at its Ultra tier described as 20x those of its Pro plan, without publishing the absolute amounts behind them.

Five times what? Twenty times what?

A multiplier looks precise because it contains a number. It is not precise when the underlying unit can move. OpenAI says, "Plus subscriptions may include usage limits such as message caps." Its Pro pricing says, "Unlimited subject to abuse guardrails." The Codex guidance says, "Usage limits vary by plan," and its pricing page adds, "Additional weekly limits may apply."

Claude is explicit that Max has two weekly usage limits plus further weekly and monthly caps. Yet it does not publish an exact message count, and consumption changes with message length, attachments, conversation history, tools, model and feature choice.

There may be good technical reasons for flexible limits. The published guidance says consumption changes with task size, context, model and tools. A short answer and a long coding job can therefore draw differently on the allowance. The problem is commercial measurement. If the denominator is not stable, a customer cannot reconstruct what 20x bought.

I started on the Istanbul exchange in 1993. We had a word for shares that traded on the next announcement instead of the last balance sheet: they were priced on hope. Some of those stories ended well. The habit of checking which ones did is the habit I am applying to my own card statement.

A promise without an inspectable counterpart in the terms has no clean audit trail. If it cannot be measured against the agreement, it cannot be checked at all.

3. The rules can change during the game

Flexible language matters because access rules and model choices do change while subscriptions are active.

Anthropic announced extra weekly limits on 28 July 2025 and put them into effect on 28 August 2025, adding them to the existing five-hour window. The company said it expected the change to affect less than 5 percent of subscribers; no independent check of that figure exists.

Cursor changed its old monthly 500-request structure to $20 of API-priced usage on 16 June 2025. On 4 July 2025, it acknowledged that its unlimited wording had not clearly explained that the term applied only to Auto, and it opened a refund route for unexpected extra usage charges.

OpenAI made GPT-5 the default automatic system on 7 August 2025, then returned GPT-4o to the model selector for paid users on 12 August 2025. Some users reported a weaker tone or task result, but nobody has produced a controlled measure that settles the question.

These are different events. They should not be forced into one story about motive. The common point is narrower. The subscriber buys access at one moment, while the provider retains substantial control over quotas, packaging and model availability after payment.

That control may be operationally necessary. It still has value to the seller and cost to the buyer. When the unit is vague and the rules can move, the buyer carries more uncertainty than the monthly price suggests.

4. The quiet exchange

None of what follows is an accusation; it is what the policy pages say. The subscription is not always a simple exchange of money for software. On some consumer accounts, the conversation can also become training material.

In personal ChatGPT Free, Plus and Pro workspaces, prompts, responses, files and other content may be used for model development, and data sharing is enabled by default. A user can turn off "Improve the model for everyone," after which ordinary new conversations are excluded. Explicit feedback can still make the related conversation eligible, while Temporary Chat remains excluded from training.

Anthropic's consumer position changed in 2025. Its 1 May policy said consumer inputs and outputs would not be used for training outside listed exceptions such as feedback or explicit participation. A new arrangement was announced on 28 August and took effect on 8 October, using opt-out policy language for Free, Pro and Max accounts. When the improvement setting is on, covered new or resumed conversations can be retained for five years, compared with an ordinary 30 days when permission is not given. The official material does not establish the starting visual position of the choice shown to existing users.

For personal Gemini Apps accounts, Keep Activity is on by default for users aged 18 or over. When it is on, chats and shared content may be used to improve services, including training generative AI models with human review. Users can turn it off, and Temporary Chats are not used for training.

Now compare the business side. OpenAI says inputs and outputs from Business, Enterprise, Edu and API products are not used for training by default. Anthropic says its commercial products also start outside model training by default. Eligible Google Workspace editions keep organisational content out of training beyond the organisation's domain without permission.

Under the consumer default the individual pays and contributes training data; under the business default the company pays and its content stays out. Defaults, not destinies, but defaults are where most people stay.

I do not need a conspiracy theory to question this design. Defaults are choices. The useful question is why the burden of changing the setting sits more often with the individual who is already paying, while the commercial customer more often begins outside training. The exchange is quiet because it lives in account settings and policy language, not because it is imaginary.

5. A fair counterweight

I am making no allegation of wrongdoing. I work from the assumption that compute is genuinely expensive. Nobody outside these companies has seen their cost accounts, so I will not invent a margin. What the published guidance does show is that task size, context, model choice, tools, features and conversation length can change quota consumption. That supports treating the resource burden as variable, without claiming to know its exact cost.

Packages and billing structures move. I expect unit prices to fall, but I cannot point to a clean historical series, so I hold that as an expectation, not a finding.

Opt-out routes are real. OpenAI, Anthropic and Google each publish a consumer control that can limit the use of future chats for model improvement, with provider-specific exceptions and retention rules.

The balance also appears in outcomes. These tools have produced work for my companies that I was glad to have. I am not publishing internal figures, and I will not dress anecdotes up as data. A peer-reviewed workplace study found measurable productivity gains in a defined customer-support workflow, while its authors limited the finding to one tool, one company and one occupation.

That is why this is not an argument for automatic cancellation. It is an argument for accounting. A fair test must be able to return a positive result. It must also be allowed to say that an impressive tool did not justify a recurring bill.

6. Counting it myself

The providers do not publish this measurement. So I will run it on my own subscriptions, in my own business, and publish what I find.

For the next month, I will record the work produced by both subscriptions. I will use repeated, real tasks where possible. I will count total time, including prompting, waiting, checking, correction and rework. I will record whether each output was accepted, lightly edited, heavily rebuilt or rejected. I will note quota interruptions, forced model changes, retries and extra spending.

It is the same method I would use on any business I run: establish a baseline, repeat the same real tasks, put the usage records next to my own judgment, and count rework as a cost, not a footnote. A full cost view includes the subscription, extra usage and human verification time; the result is then set against time saved, outside spending avoided and accepted quality.

At the end of the month, I will publish the result whatever it says. If the subscriptions save more than they cost, I will say so. If one earns its place and the other does not, I will say that. If both fail the test, I will say that too.

The reader can run a smaller test with three questions:

1. What repeated task did this subscription complete to an acceptable standard?
2. How much end-to-end time did it save after checking and correction?
3. Would I renew at the same price if no better model were promised next month?

Hope is allowed. It is not a receipt.