The Linux Foundation's Tokenomics Foundation has introduced the Big-T Notation framework, developed by Adobe cloud solutions architect Den Neff. This system is designed to evaluate and manage token expenditures in LLM and agentic systems, similar to how Big-O notation is used in algorithm theory.
The framework classifies workloads based on how token consumption grows as usage, task complexity, or agent autonomy increases. It defines a 'ladder' of complexity classes: T(1) for constant consumption (e.g., cache), T(log n) for sublinear growth (e.g., RAG with SQL filtering), and T(n) for linear growth, where tokens grow proportionally to the number of requests.
Higher complexity classes include T(n·k) for multiplicative growth, where 'thinking' models or context re-reading may increase costs by 10–20 times. T(n·k·a) represents agentic growth, where the depth of the agent tree (a) acts as a multiplier. Finally, T(∞) represents unbounded growth in production systems without cycle interruption.
Using Yandex Cloud AI Studio as an example, the author illustrates how a single user request can hide complex internal processes. A typical agentic call involves multiple steps: sending the full tool catalog, the LLM deciding which tools to use, and a second pass through the model to interpret results, which at least doubles the same request's token usage.
Analysis of Yandex's pricing shows that output tokens are generally the most expensive, while cached and tool tokens are the cheapest. The price difference between different Yandex models for 1,000 tokens can reach up to 48 times, emphasizing the importance of choosing the model based on the complexity class of the workload.
