Token list prices are a poor guide to your AI bill
In brief
According to Journal du Net, official OpenAI and Anthropic prices per million tokens poorly predict the cost of a business task. Three model pairs in the same price band were compared on four verifiable jobs. On a bug-fixing test with caching, Claude Haiku 5.5 cost 8.4 times more than GPT-6 Luna, and Claude Sonnet 5.5 1.7 times more than GPT-6.1 Sol, despite matching list prices.
Source : Journal du Net - « Anthropic vs OpenAI : l'IA la plus chère n'est pas celle que l'on croit »
What it changes for an SME
For an SME of 10 to 250 people, relying on the public price per million tokens can mislead your AI budget. According to Journal du Net, on a bug-fixing test Anthropic models counted 54% more input tokens than OpenAI models on the same code, Claude Haiku 5.5 failed all three runs, and it is billed five times more above 100,000 tokens. GPT-6 Astra still costs more than Claude Opus 5.5, but the real gap (1.7 times) is smaller than the list-price gap. Before you automate a process, measure cost and success on your own files.
What you can do
- Run the same real case, such as an invoice or a customer file, on two models in the same band and compare cost and success.
- Check how many input tokens your documents use and whether volume thresholds apply, such as the 100,000-token point where Haiku 5.5 is billed five times more.
- Turn on prompt caching and compute the cost over several passes, as Journal du Net did to match production use.
Let's talk about the time you could save.
In 20 minutes, we find where AI could save you time or money. Free, no commitment.