As more enterprises move from experimenting with AI to actually implementing AI on the mainframe, a practical question is becoming increasingly important:
How many tokens are mainframe professionals actually using and what does that usage cost?
There is plenty of discussion about LLM pricing, context windows, model performance and the future cost of AI.
But we wanted to look at something more concrete:
Real world usage today
At Geniez AI, we analyzed token usage across 300 users using the Geniez AI Framework, including users at large enterprise organizations.
The results give us an interesting view into how AI is being used by mainframe professionals today.
Across the users we analyzed, the average usage is approximately:
4 million tokens per user per month*
This is not a theoretical estimate or a benchmark generated in a lab.
It is actual usage from people using AI to work with their mainframe environments.
As organizations move from AI demos and proofs of concept to everyday usage, token consumption becomes a measurable operational metric just like CPU, storage or network utilization.
* Average across 300 daily users using the Geniez AI framework
Looking at model usage, Claude Opus 4.8 is currently the most frequently used model across customers and environments (August 2026).
But there is no single model strategy.
In fact, what we see in most enterprise environments is the opposite.
Customers are using multiple frontier models for different tasks.
The frontier models being used include Claude, ChatGPT and Gemini, alongside other specialized and emerging models.
This multi-model approach is particularly interesting on the mainframe because different tasks have different requirements.
One workload might prioritize reasoning quality.
Another might prioritize security.
Another might require the data to remain completely on-premises.
And another might simply need the lowest possible cost.
The cost of tokens also varies significantly between customers and models.
Based on what we see, the effective price can range from $0 per million tokens for example, when usage is covered by an existing enterprise agreement to approximately $60 per million tokens for some of the more expensive models.
Specialized models such as Mythos, GPT-5.6 and similar frontier or specialized models can be at the higher end of the pricing spectrum.
But looking across the usage we analyzed, the average effective cost is approximately $5 per million tokens*.
* Average effective cost across paid usage; zero-cost usage covered by existing enterprise agreements excluded
Let's take a relatively simple example. Assume an organization has 50 mainframe professionals and each uses the current average of 4 million tokens per month.
That means approximately: 200 million tokens per month
The average effective cost of 1M tokens is $5.
the token cost is: 200 * $5 = $1,000 per month for all of the mainframe team actively using AI and being more productive.
It means cost per mainframe professional per month is $20.
The cost of the LLM itself is surprisingly small compared with the value of the work being performed.
If AI saves a mainframe professional even a small amount of time each week, finding information, analyzing system behavior, troubleshooting, investigating configuration, understanding logs, writing automation or working with existing mainframe data.
The economics can become very compelling.
The takeaway? AI on the mainframe doesn't have to be expensive.
Based on the usage we see across 300 daily users, the LLM cost is approximately $20 per mainframe professional per month.
We hope these real-world numbers help organizations think more practically about implementing AI on the mainframe and, most importantly, help more mainframe professionals start using it.
