Tool that makes AI talk like the Incredible Hulk promises to cut companies' costs by 75%: 'Why use many token when few token do trick?'

“Think less “you’re right to push back, I was wrong,” and more “Hulk smash,” the outlet quipped.

The reason, unsurprisingly, comes down to money. When companies use AI for tasks ranging from coding and engineering to simply converting PDFs, inputs (the prompt you give an LLM) and outputs (the LLM’s answer) all cost tokens. Some estimates suggest that 100 words equals around 135 tokens. And those tokens aren’t free.

Depending on the needs of the company and their LLM use, spending management platform Ramp averaged out the monthly cost of AI tokens at $140,842. Over time, however, costs can run into the tens of millions.

Enter Caveman — a plugin for LLMs that touts “Fewer words. Same work” as it simplifies output responses to Hulk-like language that also resembles, as the name suggests, our earliest ancestors. Its creator, Julius Brussee, told 404 Media his plugin “cut output tokens by roughly 65–75%,” echoing the plug-in’s own Office-inspired motto of “Why use many token when few token do trick?”

The increased cost of AI usage

The Caveman plugin represents the antithesis of the once-vaunted tokenmaxxing ethos, as companies like Anthropic’s Claude and Github undergo a large-scale shift in AI pricing models to token-usage billing.

As a result, companies from Microsoft to Uber are pulling back on their AI usage due to excessive costs. One unnamed workplace, for example, unknowingly burned through $500 million in a month after letting their employees loose with Claude AI.

404 Media obtained an internal memo from digital infrastructure building company Legrand that included promoting “caveman skill” as part of their plan to reduce AI costs. They also quoted Brussee as claiming that he’s heard of developers at companies like OpenAI, NVIDIA and GitHub who are using or testing his Caveman tool.

Moneywise reached out to multiple companies, including Legrand, OpenAI, Nvidia and Caveman for comment but didn’t hear back by deadline.

Other companies, from Amazon to Accenture and Meta, are reportedly either reducing their AI usage or pulling back on internal mechanisms meant to encourage it.

Brussee, for his part, told 404 Media that he created Caveman in April after using Claude Code and noticing “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.”

Claude users on Reddit have also been onto this concept for months, with one suggesting this caveman-speak AI prompt: “Protocol: Caveman. Speak primitive. Use nouns and verbs. No grammar filler (the, is, are, of). Keep words short. Save tokens. Be blunt.”

Meanwhile, a quick ChatGPT prompt for examples of caveman-style responses shows how many words, and tokens, such brevity can save.

When asked why the sky is blue, ChatGPT offered a 170-word explanation beginning with “The sky appears blue because of a phenomenon called Rayleigh scattering.”

When asked to answer the same question with a caveman-style response, ChatGPT returned 41 words that began: “Sun make all colors. Blue light tiny wave. Tiny air bits hit blue most. Blue light bounce everywhere. Eye see blue. Sky look blue.”

Different tests using caveman-speak have yielded different results, with token savings varying depending on the prompt and the LLM. Elasticsearch Labs, for one, reportedly “measured 63.6% average [token] reduction across eight Elasticsearch scenarios with zero accuracy loss,” while another test “found roughly 45% output savings and approximately 39% cost reduction.”

Do caveman answers result in caveman intellect?

While the irony of asking advanced computer intelligence tools to speak like a caveman is evident, some do worry that such prompts could, over time, degrade the quality of the information the LLMs provide.

“You can’t help but laugh that after years of campaigning on the human-like intelligence of these chat bots, bean counting is forcing a frontal lobotomy,” gaming and tech news website Kotaku contended.

Yet studies published within the last year offer some hope. One found “LLMs tend to overthink simple problems, generating unnecessarily long outputs,” while the length of responses “can be significantly reduced while maintaining acceptable accuracy.”