Everywhere I look right now I find the same tip: tell your AI assistant to talk like a caveman. No pleasantries, no subordinate clauses, no articles. “File broken. Line 12. Fix: bracket.” For Claude Code it comes as a ready-made extension with the motto why use many token when few do trick, and it collected more than 50,000 stars on GitHub in two weeks.1
The reason is money. You pay for every word the machine writes, so you let it write fewer words. Sounds sensible. It also sounds harmless, because people have always saved on words the moment words had a price.
ARRIVING TUESDAY STOP
In the 19th century you paid for a telegram by the word. So people dropped articles, verbs and every courtesy and wrote “ARRIVING TUESDAY 2PM TRAIN STOP”. Classified ads were priced by the line, and that gave us the “2BR apt, w/d, nr subway, avail now”.
None of it did any harm. Whoever sent a telegram in the morning went on talking in full sentences at the kitchen table that evening. You read telegram style a few times a year, and it vanished along with the telegram.
Shortened language stays harmless as long as it stays in its niche. What happens when you read it every day?
The u in colour
There is an answer to that, and you see it in every American text. Americans write color, the British write colour. You often hear that American newspapers charged by the letter, so everyone dropped the silent u. There is no evidence for that.2 The truth is better.
One man took out the u: Noah Webster. From 1783 his spelling book taught American children to read, and it sold tens of millions of copies. In 1828 his dictionary followed and turned colour into color, centre into center, defence into defense. Webster wanted the young nation to stop sounding like London, even on paper.
The children read color every day, in every schoolbook. At some point nobody asked any more whether it was a misprint. The new form stuck, to this day, because a whole generation had it in front of their eyes daily.
And what do you read every day, for hours? I read the answers of my AI tools. If they talk like cavemen, I read caveman, and more of it than any telegram anyone ever received. Which leaves the question whether that matters, as long as I understand what is meant.
Newspeak
George Orwell thought that question through in 1984. The Party invents a language, Newspeak, and every edition of its dictionary is thinner than the last. Why keep cut when you have knife? From now on knife is noun and verb at once. Think goes, thought will do. You can understand everything that is left. That was never the Party’s concern.
At the end of the novel Orwell adds an appendix that has nothing to do with the plot. In the dry tone of a grammar book he explains how Newspeak works, and what it is for:
“to make all other modes of thought impossible”
Then he tries to translate the opening of the Declaration of Independence into Newspeak, the passage on the rights to life, liberty and the pursuit of happiness. It cannot be done. A single word remains: crimethink.3 Nobody can say the sentence a whole country was founded on in that language any more. So soon nobody thinks it either.
That is a novel. Can you watch something like it happen in real people?
The wine-dark sea
You can, and the oldest example is almost three thousand years old. Homer calls the sea in the Iliad “wine-dark”. A blue sea never appears, as the British politician William Gladstone noticed in 1858.4 The Greeks were not colour-blind. They sorted colours into words differently.
Today you can measure what that does. Russian has two basic words for blue, goluboy for light blue and siniy for dark blue. Native Russian speakers tell apart two shades that fall on either side of that line faster than English speakers do.5 The word gives them a handle to grab the difference by.
The same holds for thoughts. With “opportunity cost” you think in one breath what would otherwise take a paragraph. Without the handle you may still have the thought. You share it less often, more slowly and worse.
You see best what gets lost in a text that is already short. Shakespeare’s Sonnet 18 opens with four lines on a summer’s day and ends with a promise. Here it is next to its caveman version:
Shakespeare, 1609
- Shall I compare thee to a summer’s day?
- Thou art more lovely and more temperate:
- Rough winds do shake the darling buds of May,
- And summer’s lease hath all too short a date;
- …
- So long as men can breathe, or eyes can see,
- So long lives this, and this gives life to thee.
Caveman, 2026
- You hot.
- Summer worse.
- Wind bad. Summer short.
- Poem stay. You stay.
The content is all there: you are better than summer, summer ends, the poem lasts. What you lost is the question at the start, the hesitation of someone looking for the right comparison, and “darling buds of May”. Shakespeare promises that these lines will keep the person alive as long as people read. He was right; you are reading them four hundred years later. Nobody will read the caveman version four hundred years from now.
The caveman mode deletes no word from the dictionary. It deletes something from the answers you read every day: the distinctions. “Risky, but with a capped downside” becomes “bad”. If we are making that trade, I at least want to know what we get for it.
What we get for it
We pay per token, roughly a syllable or a short word. We pay for everything the machine reads, and more for everything it writes: output costs five to eight times as much as input at every major provider.6 A team that lets agents do real work reaches ten million output tokens a month quickly.
List prices from the providers, as of 18 November 2026.6
Between the most expensive and the cheapest model lies a factor of 360. On the expensive models every “Happy to help with that!” costs real money. If the caveman cuts two thirds of it, as his inventor promises, you save around 1,200 dollars a month on GPT-5.5 Pro. That sounds like a good trade. Does it deliver?
The caveman, tested
The inventor measured ten tasks and found 65 percent less output. Independent tests find less.
- output, according to the inventor
- −65%
- output in MakeUseOf’s long test
- −52%
- cost across 60 benchmark tasks
- −25%
In a typical coding session output makes up only 10 to 30 percent of the bill; input eats the rest. And the machine rereads the caveman rules on every turn, which costs tokens too. For a whole session, estimates land at four to five percent savings.1
The more interesting part is where the savings come from. Across 60 real programming tasks, Claude in caveman mode took 17 percent fewer steps and made almost 20 percent fewer tool calls.7 The machine looked things up less and checked less. The short sentences bring the smaller part of the savings.
That fits what we know about language models: they think in tokens. Let a model reason step by step first, and it solves maths problems over 90 percent of the time instead of around 60. Researchers at Zoom cut that reasoning down to five words per step. On simple questions about sport the model stayed good and needed a thirteenth of the tokens. On maths it lost four points.8
So this is the trade. You give up the distinctions in your answers, the machine gives up part of its thinking, and the bill drops by five percent.
At least it is your decision, and you can reverse it tomorrow. But what if you are not the one deciding?
Please and thank you
In April 2025 someone on X asked how much electricity OpenAI loses because people say “please” and “thank you” to ChatGPT. Sam Altman replied: “tens of millions of dollars well spent – you never know”.9 Tens of millions of dollars for courtesy, and the boss calls it money well spent.
For now. Those millions sit in a spreadsheet, and spreadsheets have people whose job is to cut them. Imagine OpenAI or another provider has to save money, and someone is told to bring down compute costs. In the instructions the model reads before every conversation, they strike the greetings first. Then the circumlocutions. Then the adjectives, the images, everything beautiful that adds nothing to the answer. Each change saves one percent, and none of them stands out on its own.
We have known since August 2025 that a provider can change the tone for hundreds of millions of people overnight. OpenAI replaced GPT-4o with GPT-5 without warning. The new model sounded cool and businesslike, “very to the point”, as an OpenAI executive later admitted. Thousands complained, a Reddit thread was titled “GPT-5 is horrible”, and a day later GPT-4o was back.10
People noticed because it happened overnight. Webster showed what happens when a change comes slowly and sits in front of you every day: at some point nobody asks. What would a tone that gets a little rougher with every update do to us?
We talk the way the machine talks
There is research on this. That people pick up the words of the person they are talking to has long been known. In 2011 Holly Branigan and her team in Edinburgh showed that we do it more with computers than with people. We adapt most when we think the computer is not very capable.11 We talk like the machine so that it understands us.
If the machine talks short and rough, you answer short and rough, for hours, every day. I don’t believe that tone stays at the keyboard. Whoever spends the day talking to their tools in the imperative takes it into the next call, the next email, to the kitchen table where people still talked in full sentences after the telegram.
Orwell needed a Party with a ministry of its own for Newspeak. Today a cost centre and one bad quarter would do for a similar result. Nobody would have to mean any harm.
What I cut
Orwell himself liked short words. In a 1946 essay he advises never to use a long word where a short one will do.12 Brevity was fine by him. What he feared was the language that erases distinctions.
I stick to that with my own tools. I strike “Happy to help!”, “Great question!” and the summary of what I just wrote myself. Adding “Be brief” to the configuration already brings a large part of the savings, according to user reports.1 The distinctions stay, even if “bad” is three tokens cheaper than “risky, but with a capped downside”.
Cut the courtesy, never the distinction.
That is my rule for my configuration. What the provider writes into theirs I don’t decide, and I don’t see. I write emails, code and texts with three or four models from three or four companies, like almost everyone I know. I can’t opt out without becoming slower than everyone else. We all hang on the same few lines.
The GPT-5 users noticed the switch because it came overnight. If it came in small steps, one percent per update, nobody would notice. Neither would I. The schoolchildren didn’t ask Webster either. One day I would write “Fine. Merge.” to a colleague and take it for my own style.
Webster needed a dictionary and a generation of schoolchildren to rebuild a language. Today one line in a file none of us ever gets to see is enough, and millions of people read the result for hours every day.
Orwell wrote his appendix on Newspeak in the past tense, looking back on a language someone had left behind. I wonder who will write our look back, and whether they will still have the words for it.
Notes and sources
-
Caveman skill by Julius Brussee on GitHub, more than 51,000 stars after two weeks (as of 1 May 2026). Independent test: MakeUseOf, −52% output and −38% overall, with input rising because of the rules. The 10–30% output share of the bill, the 4–5% overall savings and the 25–40% less output from “Be brief.” come from community reports, not controlled studies. ↩ ↩2 ↩3
-
Noah Webster, A Grammatical Institute of the English Language (1783, the “Blue-Backed Speller”) and An American Dictionary of the English Language (1828). I found no evidence for the story about newspapers charging by the letter. Later reform attempts did happen: Andrew Carnegie funded the Simplified Spelling Board in 1906, and the Chicago Tribune printed forms like thru and tho from 1934 to 1975. Almost none of it stuck. ↩
-
George Orwell, Nineteen Eighty-Four (1949), appendix “The Principles of Newspeak”. It also contains the model sentence “Oldthinkers unbellyfeel Ingsoc”, which Orwell renders as: those whose ideas were formed before the Revolution cannot have a full emotional understanding of the principles of English Socialism. ↩
-
William Gladstone, Studies on Homer and the Homeric Age (1858). Told well by Guy Deutscher, Through the Language Glass (2010). ↩
-
Jonathan Winawer et al., Russian blues reveal effects of language on color discrimination, PNAS 2007. The study supports the weak form of the Sapir-Whorf hypothesis (language shapes perception), not the strong one (no word, no thought). ↩
-
Standard prices per million tokens, retrieved 18 November 2026: OpenAI (GPT-5.5 Pro $30 input / $180 output), Anthropic (Fable 5.2 $8 / $40, Opus 5.5 $4 / $20, Sonnet 5.5 $2 / $10, Haiku 5.5 $0.10 / $0.50), Google (Gemini 3.8 Flash $0.75 / $3.75 until year end, double after). ↩ ↩2
-
Caveman Saves Tokens by Doing Less, Not Just Saying Less, 60 tasks from SWE-bench Verified, Terminal-Bench 2 and USACO. With Codex, tokens fell by only 3.9 percent. ↩
-
Silei Xu et al., Chain of Draft: Thinking Faster by Writing Less, arXiv 2502.18600 (2025). Sport: 189 → 14 tokens, accuracy 93.2 → 97.3% (Claude 3.5 Sonnet). Maths (GSM8K, GPT-4o): 95.1 → 91.1%. ↩
-
Sam Altman on X, 16 April 2025, reported by Yahoo Finance/Benzinga among others. The figure is an offhand reply, not an audited number. ↩
-
Technology.org, 11 August 2025. Altman at the time: “Ok, we hear you all on 4o.” The phrase “very to the point” comes from OpenAI executive Nick Turley. In fairness, GPT-4o had the opposite problem before: it flattered too much. ↩
-
Holly Branigan et al., The role of beliefs in lexical alignment: Evidence from dialogs with humans and computers, Cognition 121 (2011). The study measures word choice, not tone; that tone carries over too is my inference. ↩
-
George Orwell, Politics and the English Language (1946), rule two of six. ↩