Why smarter AI models can cost more despite lower token prices

Developers using advanced AI models from OpenAI and Anthropic have noticed rising costs even as per-token prices fall, and the explanation lies in hidden 'reasoning tokens.' Both companies bill internal reasoning work as output tokens, even though this chain-of-thought processing is never shown to the user. A single visible 200-token response can carry a billed output of 12,000 tokens or more, depending on how much internal reasoning the model performs. Newer models also tend to reason by default, meaning an unchanged prompt can trigger significantly more token usage than it did on an older model. Both vendors expose reasoning token counts via API fields, allowing developers to audit actual usage and identify where vague or ambiguous prompts are driving up costs.