July 6, 2026
Token of Betrayal
Price per 1M tokens is meaningless
That “cheap” AI price tag? Commenters say it’s giving fake sale energy
TLDR: The article says AI list prices can fool you: the “cheapest” option on paper may cost more to get real tasks done because companies count and use text differently. Commenters agreed the sticker price is shaky, but argued over how shaky—some called it merely misleading, while others said the whole pricing game is broken.
The big claim in Jan Iłowski’s post is deliciously blunt: judging artificial intelligence tools by price per million tokens is basically shopping by sticker price and ignoring the final bill. Different companies break text into different-sized chunks, and some models burn through way more hidden “thinking” before answering, so the cheapest-looking option can end up costing more to finish real work. That’s the plot twist. In the benchmark table, one supposedly pricier model actually finished tasks for less money than a cheaper rival, while some bargain-bin stars from Chinese labs stayed impressively low-cost but didn’t always scale in the simple way the headline prices suggest.
But the real action is in the comments, where the crowd instantly turned this into a “well, actually” cage match. One camp pushed back on the article’s title, insisting price per token isn’t meaningless so much as wildly misleading. Another commenter dropped a sneaky little bombshell: on their setup, turning thinking higher made tasks both faster and cheaper. Yes, the internet’s favorite twist: spend more brainpower, save money. People also dragged in the overlooked villains of the bill—caching discounts, subscription pricing weirdness, and models that are just plain too chatty.
The hottest mood? Nobody trusts the list price anymore. One commenter basically called per-token pricing a clown show if monthly plans charge totally different effective rates. Another skipped the drama and pitched a hybrid DIY system to cut costs altogether. The overall vibe: the community isn’t just comparing prices—they’re side-eyeing the entire menu.
Key Points
- •The article argues that listed price per 1M tokens is not a reliable standalone metric for comparing AI model costs.
- •It states that tokenizer differences between providers and even between models from the same provider can significantly change billable token counts.
- •The article says hidden or obscured "thinking" tokens can dominate usage costs and vary widely across models.
- •A comparison table combines model list prices, Artificial Analysis benchmark scores, and cost per benchmark task for several American and Chinese models.
- •The article highlights cases where a model with higher nominal token pricing, such as GPT-5.5 xhigh, can still be cheaper per completed benchmark task than a lower-priced alternative like Claude Opus 4.8 max.