A Team Analyzed 110 Videos in Claude Code: 95–99% of Tokens Go Toward Rereading

Claude Code limits have long been a topic of their own among vibe coders: everyone feels their weekly allowance melting away, but few know exactly where it goes. The team behind the Tekhnokafe channel decided to find out. It produced 110 videos with Claude Code and broke down the costs line by line, like an itemized receipt.

How it all started

The trigger was a video by a blogger who said a ten-minute episode had cost him about $2 and 4% of his weekly limit. The channel’s creators recalculated their own figures and found a much more modest picture.

A typical video consumes about 0.7% of their weekly allowance on the Max plan. The added voice-over costs 8 cents. The most expensive entry in their log is 10.2%, but the creators openly warn that this is an estimate: in August, they were still debugging the pipeline itself, and some of the spending went toward experiments. The most expensive precise measurement is 5.1%. The range from 0.7% to 5.1% speaks for itself: a video’s cost depends not on its length, but on how the workflow is organized.

Where the tokens actually burn

The main finding is not about video, but about how models work. Between 95% and 99% of tokens are spent not on new text or code, but on rereading. On every turn, the model receives the entire previous conversation again. The longer the session, the more expensive every subsequent exchange becomes—even if you are merely asking it to change a single word.

A simple example: you are five hours into one session that has included everything in sequence—the script, revisions, script debugging and frame analysis. A request to “make the headline shorter” drags that entire tail along with it. The creators tested this in practice. The same style cost 78 million tokens in a bloated session and 13 million in a fresh one: a sixfold difference for the same result.

What they changed in the process

The conclusion was practical and easily transferable to any agent-based project.

First, every new task goes into a new conversation. Once the script is finished, open a clean session for the voice-over. Pass along the necessary context as a short summary or a file, rather than the entire history.

Second, a cheaper model handles routine work. Renaming files, compiling a list or converting content to a template does not require the most powerful model. The expensive model is reserved for decisions where quality matters.

Third, frames are reviewed as a sheet rather than one by one. If the model examines images sequentially, every request once again carries the entire context with it. A contact sheet containing dozens of frames in a single image requires one pass instead of many.

Limitations and conclusion

Some caveats up front. This is one team’s experience on one plan and with one task: video production. Some figures, as the creators themselves acknowledge, are estimates. A percentage of the weekly limit is not a universal unit either: Max, Pro and API plans have different quotas, and Anthropic may change its accounting rules. You cannot directly compare your bill with someone else’s.

But the underlying logic carries over. If you are developing with an agent rather than merely writing prompts, the main cost lies in context length, not request complexity. It therefore helps to begin with simple habits: watch how bloated a session has become, do not drag old context into a new task, and do not spend a powerful model on mechanical work.

For product teams and founders embedding AI into their processes, the conclusion is even more practical. The cost of an operation should be calculated not by the price of a single generation, but by the cost of the entire session. One good video consuming 0.7% of a weekly allowance and the same video consuming five percent differ only in the discipline applied to context management. A useful starting point is the question posed by the creators: which line item on your Claude bill is the largest? Most likely, the answer is rereading.