Politeness is free… unless you’re an AI company footing the cloud bill.
Over the weekend OpenAI chief Sam Altman admitted that tiny courtesies like “please”, “thanks”, or a friendly “you’re welcome” are running up “tens of millions of dollars” in extra compute costs for ChatGPT every year.
Below is a plain‑English look at why good manners are so expensive, how the maths stacks up, and whether you should feel guilty about being nice to a chatbot.
Where the “millions” figure comes from
* A user on X wondered how much electricity OpenAI wastes when people add polite words to their prompts.
* Altman replied: “Tens of millions of dollars – well spent, you never know.”
* He was talking about operating costs (GPU time, data‑centre power and cooling), not a line item on your subscription fee.
Why two extra words add up fast
-
Everything you type is a token.
Large‑language models break text into small pieces called tokens (roughly ¾ of a word). More tokens = more maths = more electricity. -
ChatGPT is enormous.
Reuters says the service passed 400 million weekly active users in February 2025. Even if only a quarter of them toss in a quick “please” or “thanks” once a day, that’s billions of courtesy tokens a month. -
Token prices are tiny but real.
OpenAI’s latest “everyday” model, GPT‑4.1 mini, charges $0.40 per million input tokens and $1.60 per million output tokens.
A back‑of‑the‑envelope bill
| assumption | value |
|---|---|
| polite overhead per exchange | 50 tokens (prompt and bot reply like “you’re welcome”) |
| polite exchanges per day | 1.5 billion |
| daily courtesy tokens | 75 billion |
| cost @ $2 / M tokens (mix of models) | ≈ $150 k per day |
| annualised | ≈ $55 m |
That order of magnitude lines up with Altman’s “tens of millions” quip. Of course the real numbers swing around as OpenAI shifts users between cheaper mini/nano models, caches requests, and squeezes more efficiency from its hardware.
Environmental footprint
Money isn’t the only meter that spins:
-
Researchers estimate that even a short, three‑word completion (“you are welcome”) can evaporate 40‑50 ml of water through server‑farm cooling.
-
Digiconomist numbers put a single ChatGPT query at roughly 0.3–3 Wh of electricity, depending on the hardware.
Multiply that by billions of “polite” exchanges and the environmental tab becomes non‑trivial.
Why openai still pays the tab
-
User trust & brand warmth – Polite prompts nudge the model toward polite answers, which keeps conversations friendly and discourages toxic outputs. Microsoft’s design team says manners in → manners out.
-
RLHF training data – Those tiny pleasantries teach the system about social norms. Stripping them could subtly change model behaviour in ways that break downstream products.
-
Competitive advantage – A chatbot that feels “human” retains users; the power bill is part of customer‑experience spend.
Should you stop being nice to the bot?
-
If you’re a paying API user trying to save pennies, trimming fluff from your prompts (and setting
max_tokenslimits) makes sense. -
If you’re chatting for free or on a flat‑rate Plus plan, your words cost OpenAI, not you. Feel free to keep the good vibes flowing.
-
If you care about climate impact, shorter prompts help, but the bigger wins come from model‑efficiency gains and cleaner data‑centre energy – areas OpenAI and its rivals are already chasing hard.
Key takeaways
-
A couple of polite words scale into multi‑million‑dollar cloud bills because of the sheer size of ChatGPT’s user base.
-
Token costs look tiny on paper (fractions of a cent) but explode when you process tens of billions of them daily.
-
OpenAI regards the spend as worthwhile for user experience and brand perception, even as it races to lower per‑token costs with cheaper mini and nano models.
-
You don’t HAVE to stop saying “please” – but now you know who’s paying for your manners.
So the next time ChatGPT helps you out, remember: your “thank you” might be the most expensive two words the company processes all day.
