AI has a pricing problem

As the subsidizingWhen AI companies temporarily absorb part of the real compute cost so customers pay less than the service costs to run. of tokensSmall chunks of text that an AI model reads and writes. A token can be a word, part of a word, punctuation, or whitespace. is coming to an end, companies are reaching for ways to bring down the cost of AI. A practice that could be increasing costs instead of lowering them is comparing models by $X per one million tokens. The assumption that picking a model with a lower cost per one million tokens should lead to a lower bill is not necessarily true.

“One million tokens” depends on who is counting

One important component of an LLMLarge language model: the kind of AI model behind tools such as ChatGPT, Claude, Gemini, and similar systems. is its tokenizerThe component that breaks text into the pieces a model actually processes. Those pieces are tokens.: the piece that takes your input text and breaks it into tokens. Tokenizers are not standardized across labs or models. The same text can therefore produce very different token countsThe number of tokens produced from a piece of text. This can change depending on the tokenizer. depending on which model is processing it. Take this body of text:

def parse_β_v2🚀(user_name: str = "José ☕", retries: int = 3) -> dict:
    path = "/api/v2/users/{user_id}/preferences?include=security,notifications"
    trace_id = "trace__A1B2C3__🚀__ff09"
    note = "naïve résumé — urgent; café_location=Malmö; status=pending-review"
    sql = "SELECT user_id, display_name FROM users WHERE user_name = ? AND status = 'active'"
    return {
        "path": path,
        "trace_id": trace_id,
        "user_name": user_name,
        "retries": retries,
        "note": note,
        "sql": sql,
    }

If we give that exact text to a GPT-5 series model, the tokenizer splits it into 160 tokens. If we give the same text to GPT-3, it gets split into 248 tokens instead. Same input, different model, different token count. (These counts were generated with the OpenAI tokenizer; try it here.) Because token counts can vary by model, dollars per one million tokens is not a stable basis for comparing LLM costs. A million tokens according to model A is not necessarily a million tokens according to model B.

A recent event where this became painfully evident was after the release of Opus 4.7. The new model used a different tokenizer that created roughly 30 percent more tokens compared to its predecessorThe previous version of the model., Opus 4.6. It essentially meant that users who wanted to use the new model would have to foot the bill for a 30 percent silent price hike.

Alone, this would not be much of an issue. If tokenization were the only difference, you could still adjust for it. There is, however, another factor to take into account.

All models are not equally “efficient”

Let’s ignore all of the above for a second and focus on how much one token is actually worth. Not the cost, but rather what you can get out of it: how many tokens does an LLM need to complete a certain task? As it turns out, this also varies widely between models. Below, you have a table of three different LLMs, two state-of-the-artAmong the strongest models currently available. models from the US and one from China. We are going to play a game. Given what you can see in the first table, I want you to decide which model you should run. You get to see the price per million tokens and a measure of the model's intelligence, in this case the AA Intelligence benchmarkA standardized test or group of tests used to compare model performance on specific tasks. result.

Congratulations if you made the right choice. It's either lucky guessing or you are picking up what we are putting down. Even though Claude Opus 4.8 max looks like the correct choice according to the dollar cost per one million tokens and benchmark scores, that is simply not the case. A million tokens with Claude Opus 4.8 max doesn't get you as far as a million tokens with GPT-5.5 xhigh. These models are not equally efficient.

What this means

Ultimately, cost per million tokens isn't a meaningful price indicator on its own. It's simply misleading, and companies have to wake up to that fact.

A cheaper token is not always cheaper. A more expensive model is not always more expensive. And a million tokens is not a unit of value. AI pricing needs to be evaluated the same way as any other infrastructure decision: by output, reliability, and total cost of usage. Not by the sticker price of a million tokens.