Enterprises Make AI Talk Like Cavemen to Cut

There is something genuinely funny about the current state of enterprise AI: companies spent the last three years being sold on large language models because they could write like humans, and now those same companies are paying developers to make them stop.

Meet Caveman. It is a small open-source plugin — the kind of thing that starts as a weekend experiment and ends up in a GitHub repo with more stars than its creator expected — that instructs AI models like Claude and OpenAI's Codex to strip their outputs down to the grammatical minimum. No preamble. No hedges. No "certainly, I'd be happy to help you with that." Just function. Just output. Just result.

According to 404 Media, companies are deliberately deploying this tool to stop burning through AI tokens at the rate they currently are. Staff at multiple companies — including, pointedly, OpenAI itself — have turned to Caveman. AI Weekly reports that developers at Nvidia and GitHub are also among its users.

The appeal is not hard to understand. Decrypt reports that forcing Claude to speak like a caveman slashes output tokens — and therefore costs — by up to 75%. Tech Yahoo puts the same figure on the table while noting that the plugin sits alongside other approaches like prompt pruning and model routing as part of a broader token cost problem enterprises are actively trying to solve.

Seventy-five percent is not a rounding error. That is a number that makes a CFO pay attention.

What a Token Actually Costs You

To understand why anyone cares about this, you need to understand how enterprise AI billing actually works. Most commercial LLM APIs charge per token — roughly, per unit of text processed. Input tokens (what you send the model) and output tokens (what it sends back) are both metered, though output tokens are typically priced higher.

The verbosity problem is real. Claude, in particular, has a well-documented tendency toward thoroughness. Ask it to fix a function and it will fix the function, explain why the function was broken, offer two alternative implementations, note a potential edge case you hadn't considered, and then wish you well. That is useful when you need it. When you're running thousands of automated code-review passes a day across a large engineering org, you are paying for every one of those words, and most of them are noise.

What Caveman does is essentially formalize a system prompt constraint that developers had been writing by hand for months: stop explaining yourself. Give me the output. Nothing else. Kotaku notes the irony — the plugin instructs Claude to "talk less like SNL's Master Thespian" — which is a precise description of a real problem.

There is also a secondary effect worth flagging. The model accuracy angle is not purely a cost story: stripping verbal padding can sometimes sharpen what the model is actually doing. Fewer tokens to generate means less surface area for drift, for confident-sounding hedges that obscure uncertainty, for the model talking itself into a wrong answer through its own preamble. Whether that holds across task types is an open question, but the directional claim is not obviously wrong.

The Deeper Economics

Strip away the novelty of a plugin called Caveman and what you're actually looking at is an enterprise cost-control problem that has been building since the AI spending boom of 2023-2024.

Companies adopted AI tooling aggressively, often without a clear-eyed accounting of what scale would actually cost. Token usage scales with deployment. A developer using Claude Code on their own machine is a rounding error. A company rolling it out to five hundred engineers, with automated pipelines, batch processing, and integration into CI/CD systems, is a different thing entirely. The bills arrived before the ROI math was settled.

The token cost math is complicated further by the fact that Caveman is not the only lever. Tech Yahoo's reporting makes the point explicitly: prompt pruning and model routing tackle parts of the cost problem that output verbosity alone doesn't address. Prompt pruning means sending leaner input — cutting context that isn't doing work. Model routing means intelligently directing simpler queries to cheaper, smaller models instead of defaulting to the flagship. Caveman handles the output side; those approaches handle the input side and the model selection side. A sophisticated enterprise cost strategy probably involves all three.

What Caveman represents, then, is the low-hanging fruit — the quick win that surfaces the bigger structural question. If you're running a 75% output token reduction to make the economics work, what does that say about how the economics were framed in the first place?

The OpenAI Wrinkle

The detail that employees at OpenAI itself are reportedly using this plugin is worth sitting with.

OpenAI sells Claude's API competitor products. OpenAI is also, presumably, one of Anthropic's largest competitors for enterprise mindshare. The fact that staff there reached for a tool that constrains Claude's verbosity rather than defaulting to a GPT-4o deployment for internal use is not a conspiracy — internal tooling decisions are messy and people use what works — but it is at minimum a data point about how practitioners evaluate these tools in practice versus how they're marketed.

It also illustrates something broader: in real enterprise environments, the choice of AI model is not purely about capability benchmarks. It's about cost at scale, workflow integration, and increasingly, whether you can control what you're paying for. Caveman is a community-built workaround to a vendor behavior — the tendency to generate more output than the task requires — that the vendors themselves could address in their products but haven't prioritized.

What Stays Unresolved

A few things the sources don't settle cleanly.

First: quality at the margin. A 75% token reduction is impressive, but the relevant question is what percentage of that 75% was actually useless versus what percentage was context, nuance, or explanation that mattered in some cases. Caveman works well for deterministic code tasks where output is verifiable. It is less obvious how it performs for tasks where the "right" answer is more ambiguous and the model's reasoning is part of the value.

Second: this is a workaround, not a fix. The underlying cost structure of frontier LLM APIs is set by the vendors. If token pricing shifts — either because competition drives it down or because compute costs change at the infrastructure level — the calculus for tools like Caveman changes too. It's solving for current pricing conditions. Whether it's solving for the right thing long-term depends on where those conditions go.

Third: the gap between what AI is being sold as and what it's being used for. The vision was systems that could reason, explain, collaborate. The reality, in a meaningful chunk of enterprise deployment, is automated code review and batch text processing where the ideal output is as short as possible. That's not a failure — it's a use case — but it's a different use case than the one in the pitch deck.

The companies reaching for Caveman aren't disillusioned with AI. They're using it seriously enough that the costs became real. That's actually a more mature relationship with the technology than the hype cycle suggested anyone was ready for.

Whether the vendors catch up to that maturity, or keep optimizing for demos, is the question worth watching.

Mike Wierzbicki covers game development, studio business, and industry labor for Buzzrag.