Ask a coding assistant to explain a one-line fix and it will usually hand you a full paragraph: a warm opener, a restatement of your question, a couple of hedges, and a cheerful offer to help further. Switch on a tool called caveman and the same fix comes back stripped to a fragment — no greeting, no articles, no goodbye. Same answer. Roughly a tenth of the words.
That terseness is now a deliberate corporate cost-control strategy. Over the course of 2026, engineering teams have worked out that the polished, chatty prose large language models produce by default is expensive. Every transition, every caveat, every "happy to help!" gets billed — by the token. So a growing number of companies are doing something that would have sounded absurd two years ago: paying to make their AI sound dumber.
The tool at the center of the trend is literally named caveman, and the reporting that surfaced its corporate spread — a June scoop by 404 Media — gave the moment a fittingly grim label: the "Tokenpocalypse." Here is what is actually happening, why the math is real, and why the headline savings deserve a hard squint.
Why every extra word carries a price tag
To understand why a company would want its AI to talk like Tarzan, start with how the bills work. Commercial models charge separately for the tokens you send in and the tokens they generate, and output is dramatically more expensive than input.
On Anthropic's published pricing, Claude Opus 4.8 costs $5 per million input tokens and $25 per million output tokens. Claude Sonnet 4.6 runs $3 and $15. Even the budget Haiku 4.5 tier charges $1 and $5. The ratio holds across Anthropic's entire lineup — output costs five times what input does — and OpenAI's GPT-5 models show the same lopsided pattern, with output several times pricier than input.
That 5-to-1 spread is the whole game. Anthropic's own worked example makes it vivid: a one-hour coding session on Opus 4.8 that reads in 50,000 tokens and writes out just 15,000 costs about 25 cents for all that input, and roughly 37 cents for the far smaller pile of output. The model's chattiness cost more than everything you fed it. Trim the output and you attack the expensive half of the invoice head-on.
Two quirks pour fuel on the fire. Anthropic's newest models use a tokenizer that produces about 30% more tokens for the same English text, so verbosity is quietly getting pricier per sentence. And courtesy is not free either: OpenAI's Sam Altman quipped in 2025 that users typing "please" and "thank you" cost the company "tens of millions of dollars" in extra compute. Tongue-in-cheek, but grounded in a real truth about scale.
The caveman fix, explained
Caveman is a "skill" — a small package of instructions — built by developer Julius Brussee in early April 2026, after he noticed how much of an AI agent's token spend was vanishing into hedging, transitions, and chatbot politeness inside long automated loops.
The rules are exactly what the name suggests. Drop the articles: a, an, the. Kill the filler and the hedging. Replace full sentences with fragments and use arrows for cause and effect. Preserve anything that genuinely matters — code, commands, URLs, error messages, numbers — and compress only the human-flavored language wrapped around them. The result, in the tool's own framing, is a model that speaks less like a polite chatbot and more like a terse tool. One documented transformation turns "You should wrap the object in useMemo, since a new reference is created on every render" into a blunt "New ref/render. useMemo it."
It ships with intensity dials — lite, full, ultra, and a wenyan mode that compresses answers into Classical Chinese for maximum density — and installs into Claude Code, OpenAI's Codex, Google's Gemini, Cursor, and more than 30 other coding agents. A companion command, /caveman-stats, reports how many tokens you actually saved. There is no telemetry and no server; once installed, it runs entirely on your own machine.
From hobby hack to corporate policy
What turned a clever prompt trick into a news story was the discovery that real companies are mandating it. 404 Media obtained an internal memo from Legrand, the electrical and data-center infrastructure company, that lists caveman as one of four "high-impact" cost practices — alongside defaulting to cheaper models and dialing down expensive reasoning settings. The memo's rationale was blunt: "since the billing system changed and the new quotas were implemented, we all need to be mindful of our usage."
Legrand is not alone, and the users are not all penny-pinchers. According to the same reporting, developers at Nvidia and GitHub have picked up caveman, and a senior OpenAI engineer, Shayne Sweeney, contributed the code that made it work with Codex — a striking cameo, given that terser output means less token revenue for OpenAI.
The backdrop is a broad, sudden clampdown on AI spending that 404 Media dubbed the Tokenpocalypse. GitHub moved its developer tools to per-token billing in April. Uber reportedly burned through its entire annual AI budget in four months. Walmart and Uber began capping employee usage; Amazon quietly deleted an internal "tokenmaxxing" leaderboard it had built to celebrate its heaviest AI users; Meta moved to curb consumption. The mood shift is whiplash-fast: the same firms that a year ago tied promotions to using AI more are now rationing it.
Consulting giant Accenture put a name to the turn. Much of the runaway spend, the firm found, came from unglamorous overuse — employees burning premium model calls to do things like convert PDFs into slide decks.
"We're hitting this inflection point where AI is becoming material to the cost structure." — Justice Kwak, agentic AI strategy lead at Accenture, speaking to TechCrunch
Spend, Kwak added, has become deeply unpredictable while leadership keeps asking whether the money is buying real value — the kind of question that turns a growth-era splurge into a line item somebody has to defend.
Do the savings actually hold up?
Here is where the skepticism earns its keep. Caveman's marketing cites output-token reductions of "up to 75%," and its own benchmarks show a 65% average drop across ten sample prompts — 1,214 tokens squeezed down to 294. Those numbers are real. They also measure the friendliest possible case.
When independent developer Kuba Guzik ran the viral prompt against actual coding tasks on Claude Sonnet and Opus — 72 runs in all — the savings collapsed to a soberer 9–21%, clustering around 14–21%. The reason is structural: code, commands, and file paths cannot be squashed into fragments, so on real engineering work there is simply less filler to cut. The eye-popping 75% figure, Guzik concluded, is accurate mainly for explanation-heavy chatbot answers — it is "measuring something most people aren't doing."
Two more asterisks matter. First, the skill is not free to run. Its instructions add roughly 1,000 to 1,500 input tokens every single turn, and one analysis found that on a one-off query, caveman can run about 10% more expensive than a plain prompt because of that overhead. The savings only turn clearly positive across multi-turn sessions, and they lean heavily on prompt caching, which drops the cost of repeated input to a tenth of the standard rate. Second, and more reassuring: brevity does not appear to cost accuracy. Across all 72 of Guzik's runs, both models returned 100% of the correct facts, and a separate analysis found terse instructions actually improved accuracy by up to 26 percentage points on some benchmarks, by cutting down on rambling, error-prone elaboration.
The uncomfortable punchline for anyone reaching for the tool: a stripped-down, six-line "just be terse" instruction beat the full caveman skill in Guzik's test, saving 21% on Opus against the skill's 9%, at a fraction of the setup. Sometimes the fanciest optimization is a single sentence.
| Setting | Reported output-token change | Where the number comes from |
|---|---|---|
| Explanation-heavy chatbot answers | Up to ~75% cut | caveman project claims |
| 10-prompt sample average | ~65% cut (range 22–87%) | caveman's own benchmark |
| Better Stack's own benchmark | ~45% output cut | Better Stack analysis |
| Real coding tasks (Sonnet & Opus, 72 runs) | 9–21% cut | Independent benchmark, K. Guzik |
| Single one-off query (net of skill overhead) | ~10% more expensive | Better Stack analysis |
When grunting pays — and when it doesn't
So who should actually make their AI talk like a caveman? The economics favor a specific profile: high-volume, long-running agent loops where a model narrates every step, and where that narration is prose rather than code. A support bot fielding thousands of chatty tickets a day, or an agent reasoning out loud across dozens of turns, is exactly where a 40-to-60% output cut compounds into serious money. A developer firing off a single question, by contrast, will likely pay a small premium for the privilege.
It also helps to keep terseness in proportion: it is the least powerful lever on the board. Anthropic's own cost guidance ranks the bigger wins differently — match the model to the task (Haiku for simple jobs, Sonnet for most production work, Opus only for genuinely hard reasoning), cache repeated context, and batch non-urgent jobs for a flat 50% discount. Caveman shaves one slice off one side of the bill; model choice can halve the whole thing.
There is a real irony threaded through all of this. The industry spent years and untold billions teaching machines to speak in fluent, friendly, human paragraphs — and the first thing cost-conscious enterprises did with that fluency was pay to switch it off. The caveman era is less a breakthrough than a correction. It is a reminder that every token has a price, that the chatty defaults were tuned for delight rather than economy, and that in 2026 the smartest thing your AI can say is often the shortest one. Ugh. Fix bug. Done.