If you work a lot with Claude, you run into limits or a growing bill. Ben van Sprundel of Ben AI shows mods in a video that should prevent most of that usage. The figures are his own, but the habits behind them you can apply tomorrow, even without installing a single mod.
What happened
In his video of 10 November 2026 he describes a series of mods for Claude Code. A mod is a small extension that listens in on what Claude does and changes something about it. The idea behind all of them: let a cheaper model or plain code do the heavy work, instead of the most expensive model.
According to the video description, the mods do this:
Have Haiku read large files and connector results first, the lightest Claude model, so the expensive model only gets a summary.
Let Sonnet choose the thinking level per question, instead of always working at the highest level.
Run every subagent on a suitable model, light for simple work, heavy for hard work.
A countdown clock for the cache, so you can see when a chat that has gone quiet becomes more expensive to continue.
A handover to a new chat when a conversation gets too long.
Shrink browser screenshots before Claude reads them.
Ben reports that in his own tests they cut costs by 42 to 75 percent, with the same result. Those figures come from one creator, on his own tasks, and have not been independently tested. He also sells a course programme, so read them as an indication and not as proof.
What this changes for you
The useful part lies less in the mods than in the lesson behind them: most usage comes from habits. As a director or IT lead you can make agreements about those before you install anything.
Start long jobs in a new chat. A conversation that keeps growing costs more per question, because Claude rereads the whole history. Take a short summary with you to the new chat.
Choose the model and thinking level per task. Rewriting an email does not call for the heaviest model on the highest setting.
Do not hand over whole exports when part of it will do. Ask it to use only the relevant columns or pages.
Treat a mod as software with access to your work. A mod sees what Claude sees, including customer and company data. Do not install it straight from a video. Read what it does first and test it on work without sensitive data.
We use two mods of this type ourselves, the cache clock and the handover to a new chat. We have not tested the others, so we make no claim about them.
How it works under the hood
Claude Code sends a signal at every event, a hook: on a new prompt, on reading a file, on a finished answer. A mod listens on such a hook and can change the action or let a second, cheaper model help in the background. In the video Ben mentions a threshold of about 2,000 tokens above which Haiku summarises a file first, and 250,000 tokens as the moment for the handover to a new chat. Those are his settings, not fixed values.
There are two caveats. A smaller model that summarises can leave out exactly the detail you needed. And every extra model call costs tokens and time itself. Whether a mod saves money overall therefore depends on your own work. Measure it on a few real tasks before you roll it out widely.
This news touches chapters 10 and 11 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.
What you can do with this
Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.
Written in response to a video by Ben AI on YouTube.
