I keep seeing people look at AI pricing pages and assume someone has made a typo.
Input tokens cost one amount. Output tokens cost another. Cached input may cost less again. At first glance that feels odd. A token is a token, surely?
Not really.
The easiest way to think about it is this: reading is cheaper than writing. AI can read a lot of your prompt in parallel. It can process the instructions, pasted document, previous conversation and system rules in large chunks. That is the input side.
Writing the answer is different. The model generates the response one token at a time. Each next word, fragment, line of code or JSON field depends on what came before it. The longer the response, the longer expensive compute is tied up producing it.
That is why output tokens usually cost more. You are not just paying for text. You are paying for generation.
This matters more than most people think
Once you move into agents, automation, Copilot Studio, Azure OpenAI, GitHub Copilot, or anything that runs repeatedly, the economics change quickly.
A short prompt that produces a long report can cost more than a large prompt that produces a tiny answer. That surprises people. They assume the big document is the expensive part. Sometimes it is not. The real cost can be the polished, verbose output they asked the model to create.
Ask an AI system to read a SharePoint policy library and return three risks, and you are probably dealing with an input-heavy, output-light workload. That can be relatively efficient.
Ask it to create a 25-page report, an executive summary, a remediation plan, a Teams post, a client email and a formatted table every time it runs, and you have created an output-heavy workload. That is where the bill starts to move.
The MSP lesson is simple: design the output
This is where MSPs need to stop treating AI as magic and start treating it as infrastructure.
When we built servers, we cared about CPU, RAM, disk and backup windows. With AI, we need to care about prompts, context, output length, caching and repeatability.
The bad habit is asking for everything every time. “Give me the full report.” “Include all the detail.” “Make it comprehensive.” That sounds harmless until the same workflow runs fifty times across fifty tenants.
A better approach is to be deliberate. Ask for the smallest useful output first. Use summaries where summaries are enough. Generate detailed reports only when there is a reason. Reuse stable instructions and context where caching is available. Put spending limits around anything consumption-based. Review what the agent writes, not just what it reads.
Inside Microsoft 365, this means being clear about the difference between ordinary Copilot use in Word, Excel, Outlook or Teams and consumption-based AI work that may be billed differently. A user drafting an email is one thing. An agent chewing through documents and producing long artefacts all day is another.
This is not a pricing trick
I do not see the input/output price split as some mysterious vendor tax. It reflects how the technology behaves.
The mistake is pretending it does not matter.
AI costs are not just about how many people have a licence. They are about what those people, agents and workflows ask the model to produce. The output is where the hidden weight often sits.
So the practical rule is this: do not just prompt for the result. Design the cost shape of the result.
That might be the difference between AI being a useful business tool and AI becoming the next cloud bill nobody wants to open.
CIAOPS 