The Hidden Price of Politeness in AI Prompts

We’ve all grown up hearing that good manners are free. A simple “please” or “thank you” doesn’t cost anything – at least not in human conversation. But when you’re speaking to a large language model like GPT-4, societal norms or niceties cease to be a currency. AI doesn’t interpret politeness; it processes it, and every word, no matter how well-intentioned, adds a little more computational work.

Sam Altman, CEO of OpenAI, recently pointed out that these extra pleasantries whilst seemingly harmless are contributing to significant infrastructure costs. The figure he calculated? Tens of millions of dollars. It sounds dramatic, but the numbers hold up.

So how expensive is a “thank you”, really? And does being polite to a machine still make sense in a world where every token has a price?

Counting Tokens

Large language models operate on tokens, chunks of text that include parts of words or punctuation.

$0.03 per 1000 input tokens, $0.06 per 1000 output tokens

Current GPT-4 charge example

A word like “please” might use one token. A slightly more padded phrase, like “Can you please tell me the capital of France? Thank you,” adds a few more tokens than a blunt prompt like “Capital of France?”

On a per-prompt basis, the cost of those extra tokens is negligible. You’re adding maybe $0.0001 to the bill; virtually nothing. But this isn’t about a single user or a single prompt, it’s about scale.

OpenAI’s systems process billions of queries each month. If even a small fraction of those includes niceties, the volume of additional tokens becomes significant. Add in the cost of output – responses like “You’re welcome” or “Of course, happy to help” – and suddenly you’re looking at hundreds of millions of extra tokens generated every day. That’s where Altman’s “tens of millions” figure starts to make sense.

Energy, Not Just Money

The financial cost is only part of the story. Every token processed requires compute power, and compute power draws electricity. Estimates suggest that each ChatGPT prompt consumes around 2.9 watt-hours of energy. That’s significantly more than a traditional search engine query, and the difference adds up fast.

Including a few extra tokens in a prompt doesn’t move the needle in isolation, but at scale these small increments accumulate. The more polite the interaction, the more the servers have to work, and the more electricity they consume.

The carbon impact, too, is measurable. A single extra token isn’t emitting much on its own, but millions of extra tokens processed daily across data centres around the world create a cumulative environmental footprint. The difference between terse prompts and wordier, more courteous ones may be minor in individual terms, but the aggregate impact is not insignificant.

Different Models, Same Story

This isn’t just an OpenAI issue. Every major AI platform – whether it’s Claude, Gemini, DeepSeek, or Mistral – follows a similar model. Tokens in, tokens out. Compute costs, energy costs, and carbon emissions all tied to how much text the system needs to handle.

Even if you’re running models locally, on your own hardware, the dynamic doesn’t change. You’re still paying the price; just through your electricity bill rather than an API invoice. And if you’re using a larger model like GPT-4, the power draw per token is higher than it would be on smaller, more efficient models like Claude Instant or Mistral 7B.

The more complex the model, the more it costs – financially and energetically – to generate even basic replies. Politeness, in this context, means longer inputs and potentially longer outputs. That translates to more work for the machine.

Do Manners Still Matter?

All of this raises a fair question: is it still worth being polite to an AI?

There’s a practical argument for yes. Prompt tone can influence the model’s behaviour. Users who are more courteous in their prompts tend to receive responses that are more cooperative and balanced. From a product design perspective, that’s not an accident, it’s a feature; the experience feels smoother when interactions are framed positively.

I put this to Filippo Sassi, Head of our AI Labs, here’s how he put it:

“Words like ‘please’ were part of the training data. The embeddings respond far better to them. When we experimented with stripping them out – treating them like stop words – the quality of the responses fell off a cliff.”

Filippo Sassi, Head of AI Labs, Version 1

There’s also the brand angle – OpenAI and other providers are incentivised to make AI feel natural, approachable, and human-like. That includes mirroring politeness, even if it adds a few extra microseconds of GPU time and a bit more server load.

Altman himself has called this token overhead “money well spent.” And that’s the trade-off: a slight cost increase in exchange for making interactions feel less transactional.

Manners: Cheap, Not Free

You don’t need to start issuing commands like a robot to save money or energy – but it’s worth knowing that, these days, even a “please” carries a price tag.

It might be fractions of a cent, but whether you’re paying for tokens or burning through compute on your own hardware, the cost lands somewhere. API users feel it directly, everyone else pays indirectly: through infrastructure strain, energy demand, and the emissions that come with it.

So, does the old phrase still work? Not quite. Not in this context.

Manners, it turns out, are still cheap. Just not entirely free. There’s an old saying that friends and good manners will carry you where money won’t go but, when it comes to interfacing with AI technology, we’ll have to make very human value judgements on the costs.

Nathan Marlor is our Head of Data and AI. For more detail on how our AI expertise is driving innovation and helping businesses around the world transform,  click here.

Other AI related blogs