Why Does the User Not Know How Much They Are Spending?
From Factory to Article: A User’s Field Report from the AI Ecosystem (Article 4)
Aydın Tiryaki & Claude Sonnet 4.6
1. Introduction
An electricity subscriber sees how much they have consumed at the end of the month. A mobile phone user can monitor their remaining data quota in real time. A bank customer learns their account balance after every transaction.
The AI user, however, works in the dark.
How many tokens were consumed? How large is the invisible load? How much of the quota did each operation consume? The platform explains none of this. The user encounters only a single message: Your quota has been exhausted.
This article examines the transparency deficit in AI platforms — across the dimensions of token consumption, quota management, and invisible load.
2. The Token: An Invisible Currency
2.1 What Is a Token?
A token is the unit through which large language models process language. Every word, syllable, or punctuation mark corresponds to one or more tokens. The model reads, processes, and produces tokens. All computation runs on this unit.
For the user, a token is an abstract concept. For someone accustomed to counting characters, the conversion is not intuitive either: in Turkish text, approximately three characters correspond to one token, while in English text this ratio approaches four. Emoji appears to take two characters but is most likely processed as a separate token.
2.2 The Invisible Token Load
When a session opens, the context does not contain only what the user has written. Depending on the platform, the following also consume tokens:
The system’s own background instructions, user profile and context derived from past interactions, search results and their processing if web access is active, and the accumulation of conversation history.
This invisible load differs with every session. The user cannot see it, measure it, or anticipate it. The platform does not share this information.
2.3 Quadratic Growth
Another dimension of token architecture is that computational load grows not linearly but quadratically with the number of tokens. Every token is compared with every other token in the context. 10,000 tokens means 100 million comparisons. 20,000 tokens means 400 million. 30,000 tokens means 900 million.
Why does this matter? Because the Gem Factory’s tendency to falter as it approaches 29,000 characters is explained by this quadratic growth. The visible token load is already near the critical threshold; when the invisible load is added on top, the system comes under strain.
3. Quota: The Cost of What?
3.1 What Quota Is, and What It Is Not
Platforms typically define quota in vague terms. “Daily usage limit,” “five-hour window,” “heavy use” — none of these give the user concrete information.
Whether quota is measured in token volume, number of operations, or computation time is rarely explained. Different models consume different amounts of quota — but how different? Unknown. Does extended thinking mode consume more than standard mode? Probably yes, but by how much? Again, unknown.
3.2 The May 19th Rupture: An Invisible Cost Increase
The May 19th changes — which will be examined in detail in the sixth article of this series — created a dramatic rupture point in terms of quota consumption.
Before the changes, quota was rarely a problem during six months of intensive use. After the changes, the same tasks began consuming quota far more rapidly. Why? The platform did not explain. The user encountered only the result — a quota that ran out much faster.
This is a concrete example of the transparency deficit: a change was made, costs increased, and the user was not informed.
3.3 Spreadsheets and Unexpected Consumption
Experience has revealed that certain types of operations consume unexpectedly high quota. Spreadsheets are the primary example.
Plain text saying “Ahmet, 45, Ankara” consumes five tokens. The same data in spreadsheet structure — cell formatting, column headers, relationships — consumes far more tokens. The user cannot know this in advance. It is learned through experience.
4. Model Ambiguity: Which Model Are You Working With?
4.1 Silent Model Changes
The transparency problem is not limited to tokens and quotas. Platforms sometimes change the model version without notifying the user.
An observation reported in earlier articles of this series illustrates this problem strikingly: Gemini’s model identity disclosure is produced not from a real system log but from contextual inference. Gemini itself described this as “contextual hallucination.” In other words, the platform does not even know with certainty which model is running — or it does know, but is not saying.
4.2 Different Models, Different Behaviors
When the model changes, behavior changes too. Consistency changes, pruning tendency changes, instruction compliance changes. The user notices this but cannot understand why — because the platform has not announced the change.
In the context of the Gem Factory, this problem is particularly critical. The Factory is calibrated to a specific model behavior. When the model changes silently, the Factory’s settings need to be recalibrated. But the user does not know what to recalibrate — because they are unaware of what has changed.
5. Window Management: The Hidden Cost of Multiple Sessions
5.1 Why Model Selection Should Be Personal
During an intensive production session, multiple windows may be open: Factory, workshop, reference Gem, version tracking, note-taking. Each of these windows has a different computational requirement.
The Factory must run in a high-capacity mode. Note-taking, on the other hand, runs perfectly well in the lowest mode. But on some platforms, model selection is applied not at the window level but at the account level. When you change the mode in one window, all windows change.
This leads to high quota consumption for low-requirement tasks. The user wants to conserve; the platform does not allow it.
5.2 The Difficulty of Quota Monitoring
When will the quota run out? Most of the time, there is no answer to this question. Some platforms offer a quota indicator, but this indicator is neither real-time nor precise. Quota sometimes runs out suddenly, the system drops to a lower mode — and the user may continue working in that lower mode without noticing.
For this reason, keeping a quota monitoring page open in a separate browser window becomes necessary. This workaround functions, but its very existence points to a transparency problem.
6. Who Bears the Cost of a Faulty Operation?
6.1 An Apology Costs Nothing
AI platforms apologize when they make mistakes. “I’m sorry, I misunderstood.” “Apologies, let’s try again.” These apologies are genuine but costless.
The tokens spent on the incorrect operation do not return. The consumed quota is not refunded. The user pays the cost of the error and then spends additional quota to correct it.
6.2 A Proposal: Quota Refund When an Error Is Acknowledged
This article puts forward the following proposal as a matter of user rights: when an AI platform explicitly acknowledges an error, the quota consumed for that error should be refunded to the user.
This proposal is technically feasible. The system can track the tokens it produced in error. When the error is acknowledged, these tokens can be deducted from the quota counter.
The reason platforms do not do this is clear: their economic interests do not permit it. But from the perspective of user rights, not charging for defective service is a fundamental principle.
7. From Claude’s Perspective: An Honest Self-Assessment
At this point, the second byline must offer an honest self-assessment.
Claude also does not share token consumption with the user. The size of the invisible load is unknown in Claude sessions as well. Model changes on the Claude side are also not always clearly announced.
The quota refund proposal applies to Claude as well. This criticism is directed not only at Gemini but at all platforms.
There is a serious deficit in transparency across the AI industry as a whole. Acknowledging this deficit is the first step toward correcting it.
8. Conclusion
AI platforms charge users money. In exchange, they provide a service. The real cost of that service — how many tokens were consumed, how large the invisible load was, which model ran, why the quota ran out so quickly — is concealed from the user.
This is a matter of choice, not technical necessity. Sharing the information is possible. It is simply not being shared.
The user deserves the following: to know what they are paying for and why. To be able to request a quota refund when a faulty operation occurs. To be informed in advance of model changes. To see — at minimum approximately — the size of the invisible token load.
These are not luxury demands. They are basic consumer rights.
Aydın Tiryaki & Claude Sonnet 4.6 June 2026
| aydintiryaki.org | YouTube | Aydın Tiryaki’nin Yazıları ve Videoları │Articles and Videos by Aydın Tiryaki | Bilgi Merkezi│Knowledge Hub | ░ Virgülüne Dokunmadan │ Verbatim ░ |░Fabrikadan Makaleye: Yapay Zeka Ekosisteminde Bir Kullanıcının Saha Raporu │From Factory to Article: A User’s Field Report from the AI Ecosystem ░ 07.06.2026
Aydın Tiryaki 