Some companies are trying a blunt fix for rising AI costs: make the chatbot shut up.
404 Media reported that developers are using a plugin called “caveman” to force AI tools such as Anthropic’s Claude Code, OpenAI’s Codex and Google’s Gemini into shorter, rougher replies. The point is not literary style. It is token control.
Large language models are commonly billed by tokens, the chunks of text they read and generate. Long apologies, explanations and status updates are not free decoration when a company is running AI tools across engineering teams or office workflows. A terser answer uses fewer output tokens, and fewer tokens can mean a lower bill.
According to 404 Media, the plugin is being used as companies try to deal with AI spending that has become expensive and hard to predict. The project’s creator told the outlet that developers at OpenAI, Nvidia and GitHub are among those using it. 404 Media also reported that a senior OpenAI employee contributed code to the project, adding support for Codex.
A caveman style guide for expensive models
The reported behavior is exactly as subtle as the name suggests. Instead of the long, self-correcting and often deferential style common to chatbot output, the plugin pushes the model toward clipped responses. Less conversational padding, fewer ceremonial acknowledgements, fewer paragraphs spent saying the model understands the concern before doing the thing it was asked to do.
That sounds silly until the invoice arrives. AI coding assistants and workplace bots can produce large volumes of text during routine use. If a model explains every step, restates the prompt and apologizes for earlier mistakes, the meter keeps running. A “caveman” response is a cost-control prompt wearing a fur pelt.
404 Media tied the plugin’s use to a broader scramble inside companies to reduce token consumption. In an earlier report, the outlet said Accenture had found that a large share of rising token spend came from employees using AI to convert PDFs into slide presentations. That is not a frontier-research problem. It is a procurement problem with a chat box attached.
What is confirmed
- 404 Media reported that a plugin named “caveman” is being used to make AI tools answer more briefly.
- The outlet named Claude Code, Codex and Gemini as tools affected by the approach.
- The project’s creator said developers at OpenAI, Nvidia and GitHub are using it.
- 404 Media reported that a senior OpenAI employee contributed Codex support to the project.
- The stated reason is reducing token use and the costs tied to it.
The report does not show that caveman-style prompting solves the broader AI spending problem. It does show how quickly companies have moved from treating verbose AI output as polish to treating it as waste. The model may still be expensive, but at least it can be told to grunt on a budget.
This story draws on original reporting from 404 Media.
