Same $5 Rate, Bill Differs by 30%, OpenAI Exec: Token Pricing Is Never Directly Comparable
Here is a summary of the article in English:
**Title: Priced at $5, Bills Vary by 30%. OpenAI Executive: Token Prices Are Not Directly Comparable**
A key takeaway from OpenAI's Codex lead, Tibo, is that a "token" is not a standardized unit for comparing AI model costs, akin to grams or kilowatt-hours. He uses an analogy: two identical pizzas priced per slice can yield different total costs depending on how they're cut. Similarly, different models use different "tokenizers" to segment text, meaning the same input text can produce vastly different token counts.
For instance, the same text was tokenized as 766 tokens by GPT-5.6 Sol and 1170 tokens by Claude Opus 5—a 34.5% difference—despite both models advertising the same input price of $5 per million tokens. This discrepancy arises because each company trains its own tokenizer based on its training data, affecting how common or rare word combinations are split.
The problem isn't cross-vendor only. Even Anthropic warns that its newer models (Claude 4.7+) use a different tokenizer, producing roughly 30% more tokens for the same text than earlier versions, so cost estimates shouldn't be reused across model generations.
Bill differences stem from four main factors: 1) Tokenizer efficiency (input token count), 2) Caching (e.g., GPT-5.6 Sol offers a much lower cache input rate), 3) Output pricing (which can outweigh input savings in agent workflows), and 4) Context length pricing tiers (e.g., GPT-5.6 Sol charges double the input rate for entire requests exceeding 272K tokens).
Tibo also addressed user reports of GPT-5.6 Sol's context window being limited in practice, sharing configuration code to manually expand it to 1 million tokens. However, he cautioned that larger windows increase the risk of hitting higher pricing tiers as longer session histories are processed repeatedly.
The article concludes that the true metric should shift from "price per million tokens" to "price per successful outcome." The most accurate way to compare costs is to run identical real-world tasks (with the same prompts, tools, and data) through different models, accounting for all variables like tokenization, caching, output length, and context pricing. Ultimately, what matters is the total cost to complete a specific job, not the nominal token price.
marsbit08/19 08:21
