OpenAI Codex lead Thibault Sottiaux — known to many as Tibo — is pushing back on the way people shop for AI models. His message: don't just look at the price per million tokens. Because tokenizers differ from model to model, the same block of English, technical, multilingual, and numeric text used 766 tokens on GPT-5.6 Sol, while Claude Opus 5 would need an estimated 1,170. That's roughly 34.5% fewer tokens on Sol.
Both models currently charge $5 per million tokens for standard API input. For output, Sol charges $30 per million, while Opus 5 charges $25. So on unit price alone, Opus 5 looks cheaper — but that's not the whole story. If you bill that same test text as output, Sol would set you back about $0.023, while Opus 5 would cost about $0.029. Sol ends up roughly 21% cheaper.
Sottiaux reached for a pizza analogy to explain. Slice one pizza into 8 pieces at $2 a slice, and the whole pie costs $16. Slice another into 16 pieces at $1.25 a slice, and you're actually paying $20. A cheaper slice doesn't mean a cheaper pizza.
His takeaway: the smarter way to compare models is to look at what a given task actually costs end-to-end — factoring in model success rates and real token usage.
https://twitter.com/thsottiaux/status/2088866513008873560