Dive Brief:
- Enterprises are leaning into a multi-AI model strategy, increasingly adopting models from two or more providers and shifting workloads between them as IT leaders seek to balance costs and quality outputs, according to the Databricks AI Index. The report assesses enterprise AI use and spend.
- While quality has reached parity for some models, costs have not, according to the report. OpenAI’s GPT-6 Luna, for example, achieved an 89.1% task success rate for $0.05 per task. Meanwhile, Anthropic’s Claude Opus 5.5 achieved the same task success rate percentage, but at a cost of $1.62 per task — a 32 times higher cost.
- CIOs should also look beyond token costs, which don’t always translate to overall price, Databricks found. While Gemini 3.8 Flash boasts a low token price, the model emerged as one of the most expensive at $11.97 per task. “Cost per task depends on both token prices and the amount of work a model must do to reach an answer,” according to the Databricks report.
Dive Insight:
Beyond token costs, model task efficiency and quality considerations, CIOs should also consider the fundamental metric ultimately affecting the bottom line — enterprise AI users.
A traditional enterprise user spends an average of $2.49 on AI per day. Meanwhile, the 99th percentile user spent $699.74, around 281 times more, according to the Databricks report. The stark difference indicates that while users are increasing how much they use AI, the largest consumption — and a majority of the AI bill — stems from only a handful of people.
“For enterprises, the trend makes visibility into the distribution of spending especially important,” the report said. “Understanding those differences helps organizations set more targeted budgets and spend controls without unnecessarily restricting usage.”
Global end-user spend on AI models and platforms is projected to reach $64 billion in 2026, an increase of 63% from 2025, prompting CIOs to more closely scrutinize AI budgets and usage efficiency, according to analyst firm Gartner.
As AI consumption bills rise and employee AI usage levels vary, it’s becoming more critical for CIOs to focus on the return on spend, Tal Saraf, corporate VP and chief technology and AI officer at Microsoft, said in a Thursday blog post.
“In some groups, it’s a very small number of people using a very large number of tokens,” Saraf said. “Knowing who is consuming the most tokens and which models they’re using is incredibly valuable, as we strive to ensure people are using the right models for the right tasks. The point isn’t to block our people’s usage, but to educate them.”
To help manage AI spend, enterprises should track tokens by app, workflow, agent and user to gain telemetry, according to the Microsoft blog. IT leaders should also set budgets and guardrails early, and educate employees on how to use AI efficiently. Finally, CIOs can turn to lessons learned from cloud FinOps and apply them to AI, which includes building budgets based on peak consumption rather than average consumption.