Tibo Reminds AI to Compare Prices Not Just Based on Token: GPT-5.6 Sol Unit Price Higher, Yet Total Bill Lower

According to Dynamic Beating monitoring, Thibault Sottiaux, the head of OpenAI Codex, pointed out that when comparing the prices of different AI models, one should not only look at the "price per million Tokens." This is because each tokenizer (the rule for splitting text into Tokens) is different. For the same batch of English, technical, multilingual, and numerical text, 766 Tokens are used on GPT-5.6 Sol, while Claude Opus 5 is estimated to require 1170 Tokens, about 34.5% more.

Currently, both models have a standard API input price of $5 per million Tokens. However, the output price is $30 for Sol and $25 for Opus 5. Looking solely at the output price, Opus 5 is cheaper. But if we bill Tibo's test text based on the output, it would cost approximately $0.023 for Sol and $0.029 for Opus 5, making Sol actually about 21% cheaper.

Tibo used a pizza analogy: one pizza cut into 8 slices, each costing $2, totaling $16; another pizza cut into 16 slices, each costing only $1.25, yet the total is $20. "Cheaper per slice" does not mean the entire pizza is cheaper.

He believes that a more reasonable way to compare is to look at "how much it costs to complete a task" and also consider the model's success rate and actual Token usage.