Companies are deliberately making their AI tools speak like cavemen in an attempt to stop burning through AI tokens and curb their massive expenditure on AI, 404 Media has found. The tool turns the usually verbose outpost of LLMs like Claude Code, Codex, or Gemini into a much more to the point answer. Think less \u201cyou\u2019re right to push back, I was wrong,\u201d and more \u201cHulk smash.\u201dUse of the caveman plugin is in direct response to the skyrocketing and unpredictable cost of AI. As 404 Media previously reported, companies are scrambling to stop spending so much on AI, with consulting giant Accenture finding much of the \u201csoaring token spend\u201d is thanks to people using AI to convert PDFs to presentations. People using caveman include developers at OpenAI, Nvidia, and GitHub, according to the tool\u2019s creator. A senior OpenAI employee has even contributed code to the project, adding support for OpenAI\u2019s Codex tool.\ud83d\udca1Do you know anything else about token spend inside companies? I would love to hear from you. Using a non-work device, you can message me securely on Signal at joseph.404 or send me an email at joseph@404media.co.\u201cI made Caveman back in early April because I was using Claude Code heavily and noticed a lot of my token spend was going to unnecessary prose: pleasantries, hedging, transitions, and chatty language that does not really matter inside an agent loop,\u201d Julius Brussee, the creator of caveman, told 404 Media.\u00a0One company using caveman is electrical and digital infrastructure giant Legrand which, ironically, has entered the data center business. An internal Legrand memo shared with 404 Media tells employees \u201csince the billing system changed and the new quotas were implemented, we all need to be mindful of our usage of AI so we don\u2019t use up our entire budget allowance too quickly.\u201d It goes on to list four things that will produce \u201chigh impact\u201d: not always using the most powerful model; not always using high reasoning settings for the LLMs; using different more appropriate models for different tasks; and finally \u201cuse \u2018caveman skill\u2019 to reduce output consumption (without impacting code).\u201dIn 404 Media\u2019s tests of caveman with Claude Code, the plugin does make the LLM\u2019s answers much more to the point. \u201cWant changes to it?\u201d the LLM asked after I told it to review some previously written code. \u201cUses official API, not scraping,\u201d the LLM added, describing how the code worked. When I double checked caveman was installed Claude outputted, \u201cAlready active. What you need?\u201dCaveman can also display what it says is the total number of tokens saved. In my case, caveman said it had saved me around 5,800 tokens, or 65 percent.\u00a0A screenshot of caveman in action.\u201cIt makes the model speak less like a polite chatbot and more like a terse tool,\u201d Brussee said. \u201cSame substance, fewer words. In my evals, Caveman cut output tokens by roughly 65\u201375 percent versus default verbose output, and still beat a normal \u2018be concise\u2019 instruction. That number varies by workflow, but the effect was clear.\u201dCaveman users can pick their level of \u201cgrunt\u201d: lite, full (which is the default setting), ultra, or Wenyan, which translates the output into classical Chinese characters (I verified this works and now have no idea what the Claude output says).\u201cThe goal was to reduce output tokens without touching the parts where exactness matters: code, commands, paths, URLs, numbers, function names, and technical details. Caveman mostly compresses the surrounding language,\u201d Brussee added.Records on GitHub show that Shayne Sweeney, director of engineering at OpenAI, has contributed code to caveman. A commit a couple of months ago says, \u201cAdd Codex plugin support.\u201dA screenshot of caveman in action.Caveman also offers a whole agent that condenses everything down to caveman language. \u201ccaveman-code shrink everything \u2014 full terminal coding agent, caveman top to bottom. ~2\u00d7 fewer tokens than Codex on identical tasks. 20+ providers \u00b7 plan mode \u00b7 autopilot goal loop \u00b7 MIT,\u201d caveman\u2019s GitHub repository says. Caveman can also be used with OpenClaw, the agentic AI tool that went massively viral earlier this year.The plugin is obviously pretty funny but comes in response to a very real problem. In April, GitHub announced it was going to start charging customers per token rather than a flat subscription fee. Uber capped employee\u2019s use of AI tools and the company\u2019s CTO says Uber blew through its entire AI budget in just four months. Walmart also capped AI tool usage. And now you have companies using caveman.<\/p>\n
\u201cI\u2019ve heard from many individual developers and engineers inside companies using or testing it, including people at OpenAI, NVIDIA, GitHub, and DEPT,\u201d Brussee said.In leaked audio obtained by 404 Media, Accenture positioned itself as the cure to this problem, even though it encouraged clients to adopt AI as quickly as possible in the first place. In that audio a senior employee said Accenture had a new opportunity with its clients \u201cto really think about token economics.\u201dLast year OpenAI CEO Sam Altman suggested that people extending their own pleasantries to LLMs, like \u201cplease\u201d and \u201cthank you,\u201d cost OpenAI tens of millions of dollars in electricity costs.Legrand, OpenAI, Nvidia, and GitHub did not respond to requests for comment on their caveman use.Caveman\u2019s GitHub repository says near the end: \u201cCaveman save you token, save you money.\u201d<\/p>\n","protected":false},"excerpt":{"rendered":"