Does the “Talk Like a Caveman” Setting Actually Save Tokens?

Every few months, a new “prompt hack” goes viral. First it was “act as an expert”. Then it was “think step by step”. Now it’s the caveman: strip your prompt down to bare words, drop all politeness, and — supposedly — watch your token bill shrink.

“Me want answer. No long text. Short only.”

It sounds like a hack that should work. Fewer words in = fewer tokens in. But the numbers tell a more interesting story — and the part everyone repeats is the part that barely matters.

How tokens actually work

Models don’t read words. They read tokens — chunks of text produced by a byte-pair encoding (BPE) tokenizer. Common words map to a single token (the, me, want — all 1 token). Longer or rarer words split into two or more (concise is 2 tokens in cl100k_base).

The catch: because English is heavily represented in most vocabularies, most common words are already one token. Your prompt is billed in words, roughly, with a penalty for unusual vocabulary.

What we measured

We ran the exact meme through OpenAI’s tokenizer (cl100k_base, the GPT-4 family encoding) — not guessing, actually counting:

• Polite version: “Could you please provide me with a concise response to my question?” — 13 tokens
• Caveman version: “Me want answer short.” — 5 tokens
• Long polite version: “I would greatly appreciate it if you could please give me a brief and to-the-point answer, as I am quite busy right now…” — 35 tokens
• Terse version: “Answer briefly. I’m busy.” — 7 tokens

So yes: going full caveman cuts input tokens by 60–80% on these examples. The meme is real on the input side.

But here’s what the meme misses

Input tokens are the cheap half. Most providers charge 3–4× more for output tokens, and output is where a terse instruction does its real work — not by shortening what you send, but by shortening what the model sends back.

Consider the same task asked two ways. A wordy prompt invites a wordy answer: “Sure, here’s a detailed explanation broken down into several points for your convenience…” That’s 33 tokens of output. A terse instruction gets you “Sure. Here’s the answer.” — 7 tokens. That’s a 79% reduction on the expensive side of the bill.

On a representative 1,000-word task at illustrative GPT-4o pricing, the combined effect is roughly a halving of cost per call — and nearly all of that comes from the shorter output, not the shorter input.

Why it feels like magic (and sometimes is)

Terse prompts work for three mechanical reasons:

1. Brevity instruction is a strong prior. Models are trained to mirror the register you use. Short input → short output, reliably.
2. Less context dilution. Filler words compete with actual instructions for attention. “Answer briefly” buried in 200 words of politeness is a weaker signal than “Answer briefly.” alone.
3. Latency and cache. Fewer tokens means a smaller KV cache and faster first token — measurable on high-volume pipelines.

The catch

Caveman prompting is a cost optimization, not a quality one. For simple, well-specified tasks it’s fine. For anything nuanced — code review, legal text, strategy, anything where context and tone carry meaning — stripping the prompt to grunts removes the very context the model needs. The result: worse answers, more retries, and a token bill that quietly goes up.

There’s also a hidden risk: lost instruction fidelity. “Me want answer short” does not encode your constraints (format? length? tone? what not to do?). A saved 8 tokens of input is a terrible trade if it costs you a retry that costs 500 tokens of output.

The bottom line

• Clipping filler words saves input tokens — real, but that’s the cheap side.
• The expensive win is asking for shorter output: “Answer in 3 bullet points. Max 50 words.” — that’s a direct, structured brevity signal.
• Keep your constraints. Drop the politeness, never the specifics.
• On high-volume pipelines, token savings compound — but measure with your own tokenizer and your own pricing, not the meme.

The caveman isn’t wrong. He’s just optimizing the wrong half of the bill. Be terse where it costs you nothing — the output side — and keep your words where they earn their tokens: the instruction itself.