As companies grapple with rapidly rising AI costs, some are turning to a plugin called Caveman that strips away polite phrases, unnecessary explanations and conversational filler from AI responses, helping reduce token usage and lower operating costs without affecting technical accuracy. The tool is designed for coding assistants such as Claude Code, Codex and Gemini, replacing lengthy replies with short, direct responses. According to Caveman creator Julius Brussee, he developed the plugin after noticing that much of his spending on Claude Code came from unnecessary conversational language. Brussee says the plugin typically cuts output tokens by 65 to 75 per cent compared with the default responses generated by large language models. In one test, it reported saving around 5,800 tokens, equivalent to roughly 65 per cent of the output. The plugin has attracted interest from developers at companies including OpenAI, Nvidia and GitHub, according to Brussee. An internal memo from Legrand reportedly advised employees to use the "caveman skill" to reduce AI consumption after changes to the company's billing system and usage quotas. Users can choose among multiple response styles, including Lite, Full, Ultra, and Wenyan. GitHub records show that Shayne Sweeney, Director of Engineering at OpenAI, contributed code adding Codex plugin support to the project.