Anthropic recently announced the release of its latest model update, Claude Opus 4.7, with a reassuring footnote: the sticker price remains unchanged. At $5 per million input tokens and $25 per million output tokens, the company is signaling stability in an increasingly competitive market. However, a closer look at the technical specifications reveals a more complex reality. While the unit price is frozen, the number of units required to complete a task has quietly expanded—a phenomenon that might be described as "token inflation."
Two primary factors drive this increased consumption. First, Anthropic has introduced an updated tokenizer, the system that breaks down natural language into the mathematical chunks the model processes. This new version is more granular, meaning the same paragraph of text now translates into roughly 1.0 to 1.35 times more tokens than it did under the previous version. Second, the model has been tuned for "deeper analysis." In the context of large language models, "thinking more" is functionally synonymous with generating more internal and external text, particularly in complex agentic workflows.
This shift highlights a growing tension in the AI industry between performance and predictability. As models become more capable, they often become more verbose in their reasoning steps to ensure accuracy. For the enterprise customer, the cost-per-token is becoming a less reliable metric for budgeting than it once was. In the quest for higher reliability and better logic, the industry is moving toward a model where intelligence is measured not just by the final answer, but by the computational volume required to reach it.
With reporting from Xataka.
Source · Xataka
