July 13, 2026
Token me maybe
The real prices of frontier models. Tokens * Price, right?
That “cheap” AI may be secretly charging more, and commenters are furious
TLDR: The article says AI pricing pages can be misleading because different systems split the same text into different numbers of billable pieces, which can quietly raise your costs. Commenters are torn between calling it a stealth price hike, praising OpenAI’s transparency, and arguing the real bill depends on even more hidden behavior.
The big bombshell here is wonderfully petty: the price tag everyone compares may be the wrong number. The article says AI companies love showing a neat “price per million tokens,” but a token is basically just a little chunk of text, and different AI systems chop up the exact same file into wildly different numbers of chunks. Translation for normal humans: two bots can claim the same sticker price, then send you very different bills. The eye-popping example had one TypeScript code file counted as 681 pieces by GPT and 1,178 by Claude’s newer system. Same file, much bigger tab.
And the comments? Oh, they went straight for the throat. One camp praised OpenAI for at least being more transparent, with one commenter cheering that OpenAI’s tokenizer actually got more efficient, not worse. Another basically summed up the rage-posting mood with, “Anthropic is screwing us again.” That’s the kind of subtle discourse we’re dealing with.
But not everyone bought the article’s full villain arc. A few cooler heads argued this still doesn’t tell the whole story, because the real cost of an AI coding helper also depends on how chatty it is, how much extra text it drags in, and how much behind-the-scenes “thinking” it does. One commenter even popped up with the nerdy but fair question: is there any upside to the new tokenizer at all? So yes, the community is split between “this is a hidden price hike,” “actually the math is more complicated,” and “please explain why the expensive scissors are worse.”
Key Points
- •The article argues that real AI model cost depends on both price per token and how many tokens the tokenizer generates from the same content.
- •The same TypeScript file is reported as 681 tokens on GPT-5.x and 1,178 tokens on Claude's newest tokenizer, a 1.73x difference.
- •Claude Opus 4.6 and Opus 4.8 are described as having the same listed $5.00 / $25.00 rates, while Opus 4.8's newer tokenizer produces roughly 29-32% more tokens on the same code samples.
- •The article says tokenizer differences are largest on code workloads such as TypeScript, while English prose shows a smaller gap of about 1.4x.
- •The methodology used 16 real fixtures and provider tools or endpoints, including Anthropic's count_tokens endpoint, OpenAI's o200k_base via tiktoken, and token-count endpoints from Google and xAI.