AI Cost Optimization: How Caveman-Style Prompts Help Enterprises Save Big

The Token Cost Problem Is Getting Worse

AI firms have been quietly shifting from flat-rate subscriptions to usage-based billing. For individual users, that’s a nudge. For enterprises running thousands of queries a day, it’s a budget crisis.

Every word an AI model generates costs tokens. Every pleasantry, every hedge, every “Great question! Let me walk you through this step by step” — all of it adds up. And for most enterprise use cases, none of that conversational padding actually matters.

The output that gets used is the answer. Not the warm-up.

Enter the Caveman Plugin

Julius Brussee noticed this problem firsthand while using Claude Code heavily. A significant portion of his token spend was going to what he called “unnecessary prose” — pleasantries, transitions, hedging language, and chatty filler that added length without adding value.

His solution was a plugin called Caveman. It strips AI responses down to their functional core. Less “Here’s a thoughtful breakdown of your options,” more “Do this. Then this.”

The results were striking. Brussee reported a 65% reduction in token usage after deploying the plugin. That’s not a marginal efficiency gain. That’s a structural cost cut.

According to 404 Media, the plugin has since spread to staff at multiple companies — including OpenAI itself.

Why This Works: The Prompt Engineering Angle

This isn’t magic. It’s prompt engineering applied to output format rather than input quality.

Most enterprise teams focus their optimization efforts on how they write prompts — being specific, providing context, structuring requests clearly. That’s correct. But fewer teams think about constraining the response format with equal discipline.

When you instruct a model to drop conversational scaffolding and return only the functional answer, you’re not degrading the quality of the information. You’re removing the performance layer that was never meant for you — it was designed to make the product feel approachable to general consumers.

Enterprise users don’t need approachable. They need accurate and fast.

The Broader Shift in AI Economics

What the Caveman story really signals is that the AI industry’s financial model is under pressure in ways that are starting to reshape product decisions at every level.

OpenAI shuttered Sora, its video generation product, reportedly burning through a million dollars a day — even while negotiating high-profile deals. That’s not an outlier. It reflects a systemic challenge: none of the major AI firms have cracked natural profitability yet, and the cost of running these services is genuinely prohibitive.

Usage-based pricing is the industry’s way of distributing that pressure downstream. Enterprises absorb more of the real cost. And that changes the calculus for every team that built workflows assuming flat-rate access.

What Enterprises Should Actually Do

If your organization is running meaningful AI workloads, token efficiency deserves a dedicated line item in your optimization strategy. Here’s where to start.

Audit your current output verbosity. Pull a sample of AI responses from your most-used workflows. Count how much of each response is actually used versus skimmed or ignored. The gap is usually larger than expected.

Constrain response format in your system prompts. You don’t need a plugin to get terse outputs. A well-written system prompt that specifies response format — bullet points only, no preamble, max 100 words — can achieve similar results without third-party dependencies.

Separate consumer-facing and internal AI interactions. Friendly, conversational AI output makes sense when a customer is reading it. It makes no sense when an internal tool is summarizing a document for a developer. Apply different prompt configurations for different contexts.

Track token usage by workflow, not just in aggregate. Most enterprise AI platforms expose token consumption data. Use it. You’ll quickly identify which workflows are generating bloated outputs and where compression has the highest ROI.

Test compression plugins like Caveman where appropriate. If your team is heavily using Claude or similar models for code review, documentation, or internal Q&A, a response-compression layer is worth evaluating. The 65% figure Brussee reported won’t apply universally, but even a 20–30% reduction at scale is significant.

The Irony Worth Sitting With

The AI industry spent years anthropomorphizing language models. The human-like quality of responses was a core selling point — proof that the technology had arrived. Now the same companies are quietly walking that back because the economics demand it.

That’s not a failure. It’s a maturation. Enterprise software has always been about function over form. AI is finally being held to the same standard.

The teams that adapt fastest — treating AI outputs as structured data rather than conversational exchanges — will carry a meaningful cost advantage into whatever pricing environment comes next.

Smarter prompting isn’t just a productivity hack anymore. It’s a financial strategy.