Survey: only 11% of 396 businesses could forecast AI spending; Microsoft finds lower-priced models cost more than higher-priced ones on 32% of 6,800+ tasks
Artificial-intelligence use is measured in tokens, but the models' use of tokens at any given task can be unpredictable
Context & Ripple Effects
A 2025 report had already shown that falling token prices did not necessarily reduce developers’ bills, because reasoning models can consume more tokens per task. A June KPMG survey similarly found that only 26% of companies had a comprehensive view of AI costs, an early warning that AI spending visibility was incomplete.
Buyers have responded by routing work to cheaper models, including Chinese offerings, putting pricing pressure on OpenAI and Anthropic, while Shopify took the opposite approach and standardized engineers on frontier models. The new survey and Microsoft task results sharpen the problem: nominal token pricing is not a reliable proxy for what a completed workload costs.
First-order effects
- The surveyed businesses face an immediate budgeting problem: only 11% can forecast AI spending, limiting their ability to set dependable usage and procurement budgets.
- Microsoft’s results require buyers to evaluate models by total cost on their own tasks, since a lower-priced model was more expensive on 32% of the tested workloads.
Second-order effects
- OpenAI, Anthropic and lower-cost rivals face procurement comparisons based on completion cost and token consumption, not headline price per token alone.
- Companies using cheaper-model routing tools must add workload-level measurement; a low-price routing decision can raise costs when a model uses substantially more tokens to finish the task.
Third-order effects
- If task-level cost variation persists, AI procurement shifts from selecting a preferred model to operating measured portfolios of models, with observability becoming part of the inference stack.
- Model competition increasingly turns on predictable cost per useful task rather than token-price deflation, giving buyers more leverage only when they can measure workload outcomes.
The trend: Enterprise AI is entering a procurement phase in which cost control depends on task-level inference economics rather than advertised token rates.