- Primary Subject: “Caveman AI” Plugin for Reducing Token Costs
- Key Update: Developers are adopting a plugin called Caveman that forces AI tools to give shorter, more direct responses in order to reduce expensive token usage.
- Status: Emerging Tool / Industry Experiment (Unverified Broad Adoption)
- Last Verified: July 6, 2026
- Quick Answer: The “Caveman” plugin strips AI responses down to short, blunt outputs to reduce token costs, with reports suggesting it can significantly cut usage in some workflows.
AI has spent years learning how to talk like a person. Now that companies have seen how much all that talking costs, some are teaching it to communicate like a caveman instead.
A plugin appropriately named Caveman is spreading among developers who want AI tools such as Claude and Codex to stop wasting expensive tokens on pleasantries, hedging, long transitions, and the kind of chatty explanations nobody asked for.
The result is a much blunter AI that uses fewer words, burns fewer tokens, and costs less to run.
Why Are Companies Making AI Talk Like a Caveman?
Because every unnecessary word can add to the cost of using AI at scale.
Caveman was created by developer Julius Brussee after he noticed that much of his spending while using Claude Code was going toward prose that had little value to the actual task.
AI assistants have been trained to sound conversational, which often means wrapping useful information in introductions, explanations, reassurances, transitions, and offers to keep helping. Caveman cuts much of that away.
A normal AI might explain a coding problem across several paragraphs, carefully describing what went wrong and how to fix it.
With the plugin enabled, the same answer can be reduced to something closer to: “New object each render. Causes rerender. Use memo.”
Brussee claims the approach cut his output token use by roughly 65%.
That does not mean every company using Caveman will suddenly cut its entire AI bill by the same amount, since output is only one part of the cost and larger workflows can involve far more than visible responses.
For reasoning models, however, shorter exchanges can also reduce some of the text carried through a longer session. It makes sense why people like the idea.
One developer discussing the story on Reddit said an AI code review produced four paragraphs because someone had left a sentence unfinished in a comment, burning through around $8 in tokens to explain what could have been said in a few words.
Is Anyone Actually Using Caveman?
Caveman has apparently travelled well beyond one developer trying to make Claude less talkative, with 404 Media reporting that Brussee said employees at companies including OpenAI, Nvidia and GitHub had used the plugin, while a senior OpenAI employee even contributed support for Codex.
That should not be confused with those companies officially adopting Caveman across their businesses, but it shows that token waste has become a serious enough problem for people inside major AI and technology companies to experiment with cutting it down.
The plugin now works with more than 30 AI coding tools and agents.
Its growing popularity also comes at a time when businesses are becoming far more conscious of how much AI use actually costs after spending years encouraging workers to use the technology wherever possible.
The public reaction has been predictably merciless. One of the most popular Reddit responses quoted The Office: “Why waste time say lot word, when few word do trick?”
Another joked that AI no longer needs to write a five-page paper for every prompt, while developers shared their own experiences of models producing paragraphs of “botsplaining” for questions that needed little more than yes or no.
Behind the jokes, there is a legitimate problem. A verbose response does not only consume more output tokens.
It can also remain in the model’s context during a longer session, adding more text for the AI to repeatedly process while forcing the human on the other side to spend time reading information they never needed.
Is Caveman AI Actually the Future?
Probably not in the literal sense, but the idea behind it could become increasingly common.
Companies do not need every AI system to speak in broken sentences to reduce costs.
They can simply demand shorter answers, use cheaper models for simpler jobs, limit unnecessary context, and stop paying premium AI systems to complete tasks that do not need them.
Caveman turns that efficiency problem into a joke, but the reason people are using it is serious.
There are also limits to the headline-grabbing 65% figure. The savings came from Brussee's own output token use and should not be treated as a guaranteed reduction in the total cost of every AI workflow.
Some Reddit users familiar with the tool pointed out that the savings can be much smaller in complex jobs where most of the expense comes from input, reasoning, tool use, or other parts of the process.
Even so, it is hard to beat for the symbolism alone. Silicon Valley spent years selling the dream of machines that could speak just like us.
Companies finally embraced them, watched the costs pile up, and decided the future of artificial intelligence was use few word, save token, boss happy.
For more like this, stick with us here at Gfinityesports.com, the best website for gaming news, reviews, features, and guides.