Enterprise AI costs can diverge sharply from model providers’ advertised token rates, making it difficult for technology leaders to forecast what adoption will cost. The Databricks AI Index says organizations are increasingly using models from multiple providers and moving workloads among them to balance output quality and expense. Its comparisons suggest that cost per completed task, actual usage patterns and the people driving consumption can all matter more than a low per-token price.
In one comparison, OpenAI’s GPT-6 Luna and Anthropic’s Claude Opus 5.5 each recorded an 89.1% task success rate. The Databricks report put the cost at $0.05 per task for GPT-6 Luna and $1.62 for Claude Opus 5.5, a 32-fold difference despite the same reported success rate. The figures illustrate why model selection based on quality alone may not reveal the financial trade-offs.
Token rates can also give an incomplete picture of relative costs. Databricks found that Gemini 3.8 Flash, despite a low token price, was among the most expensive models in its comparison at $11.97 per task. The report attributes that gap to the work required to produce an answer as well as the price of the tokens consumed. “Cost per task depends on both token prices and the amount of work a model must do to reach an answer,” it said.
Usage patterns shape the bill
Model pricing is only one part of enterprise AI spending. The distribution of usage among employees can be just as consequential, according to the report. A typical enterprise user spends an average of $2.49 a day on AI, while a user at the 99th percentile spends $699.74 per day—about 281 times as much. That disparity suggests that a small number of heavy users can account for a substantial share of consumption.
For finance and IT teams, an organization-wide average may therefore conceal where spending is concentrated. Databricks says visibility into differences among users can help companies establish more targeted budgets and controls without unnecessarily limiting access for employees who use AI less intensively.
The issue is gaining urgency as overall spending grows. Gartner projects global end-user spending on AI models and platforms will reach $64 billion in 2026, up 63% from 2025. That forecast adds pressure on CIOs to assess not just how much AI is being used, but whether the resulting work justifies its cost.
Tal Saraf, Microsoft’s corporate vice president and chief technology and AI officer, made a similar point in a Thursday blog post. He said that in some groups a small number of people consume a large number of tokens. Knowing who uses the most tokens and which models they use can help organizations direct employees toward models suited to their tasks, Saraf said. The goal, he added, is to educate employees rather than block usage.
Build AI cost controls around real workloads
Microsoft’s guidance calls for tracking token consumption by application, workflow, agent and user. That level of telemetry can help organizations identify costly use cases and see which models are being applied to particular tasks. Without those details, leaders may see the total bill but lack the information needed to explain or manage it.
Enterprises can also set budgets and guardrails early, while explaining efficient AI use to employees. The emphasis is on matching models to workloads and improving visibility, rather than relying on a single price metric or imposing blanket restrictions.
Cloud FinOps practices offer another point of reference, according to the source material. One lesson is to build budgets around peak consumption rather than average consumption. Because AI use can vary widely between employees and workloads, an average-based forecast may not account for the people or periods responsible for the highest costs.
For CIOs, the practical cost-management challenge is thus broader than comparing token prices. They need to evaluate the cost and quality of completed tasks, monitor consumption at a useful level of detail, and understand who is using which models. As enterprise AI adoption and bills rise, those measures can help organizations manage spend while preserving useful access.