In brief: Rising model prices (e.g. Anthropic’s "Claude Fable 5" at double the cost of "Claude Opus 4.8") and inconsistent billing models are forcing companies to allocate AI spending granularly to applications and benefits rather than merely measuring usage frequency.
Rising model prices and growing token consumption are driving up AI costs at companies, while many lack an overview of spending per application. For CTOs, the question of systematic cost control is increasingly moving to the forefront alongside the sheer performance of the models.
Prices for high-performance AI models are rising noticeably. Anthropic’s new model “Claude Fable 5” costs around 9 euros per million input tokens and about 45 euros per million output tokens – twice as much as the previous flagship model “Claude Opus 4.8”. Since token-based billing is not standard across all AI applications – many companies purchase services such as ChatGPT or Microsoft Copilot via user licenses with flat-rate pricing – an inconsistent cost picture emerges. As a result, many companies lack the ability to attribute spending to specific applications.
At the same time, investment in AI continues to grow. According to an analysis of spending in Europe, OpenAI remains the leading tech provider in Germany and continues to gain ground. Anthropic, however, is catching up: average spending per customer there rose 33.1 percent year-on-year. For CTOs, this means the strategic question is shifting from “whether” AI is deployed to “how” it can be operated economically – with growing pressure from finance departments to regulate spending.
Cost drivers are manifold: poorly formulated prompts that consume unnecessarily many tokens, insufficient compatibility between applications, duplicate tool purchases, and a lack of data exchange between systems. The use of shadow AI is particularly critical, when employees use their own unapproved software instead of company-provided models. According to Bitkom, 61 percent of companies that use AI deploy it to improve their own business operations – but the technical infrastructure must be able to meet these demands.
Usage frequency alone is not sufficient for reliable performance measurement: two teams with similar usage duration may deploy the technology in completely different ways – for example in core business operations versus peripheral projects – with correspondingly different value for the company. Finance departments should therefore attribute spending specifically to individual departments and users and compare it against the expected benefit, for instance using concrete metrics such as time to first campaign draft. Only in this way can it be determined which departments are actually becoming more productive through the use of AI.