I am not asking for proprietary infrastructure details or exact financial costs. Normalised ratios and anonymised distributions would be sufficient.
When OpenAI changes ChatGPT token limits, message limits, model quotas, or other usage policies, could it also provide some anonymised aggregate usage statistics explaining the workload behind those decisions?
I am not asking OpenAI to disclose commercially sensitive information such as exact infrastructure costs, GPU counts, data-centre utilisation, or proprietary model architecture.
However, if a policy change is justified by capacity, computational cost, fairness, or maintaining service quality, it would be useful for users to understand what actual ChatGPT usage looks like.
For example, OpenAI could publish statistics such as:
- What is the average conversation depth?
- What are the median (P50), P90, P95, and P99 conversation depths?
- How many conversations does an average user start per day?
- How many user turns and total messages does an average conversation contain?
- What are the average input-token and output-token counts per conversation or per active user?
- What proportion of total inference workload or cost comes from long conversations?
- How do usage distributions differ between Free, Plus, and Pro users?
- What is the relative computational cost of reasoning models compared with standard models?
- How much additional workload is generated by tools such as Web Search, file analysis, image generation, Agents, and Codex?
- How much computational cost is reduced through context caching / prefix caching?
It would also be helpful to show distributions rather than averages alone.
For example, an average conversation depth of 100 tells us very little if most users have a depth of 10 while a very small number of heavy users regularly reach 1,000 or more.
Publishing the median, P90, P95, P99, and perhaps P99.9 would make it much easier to understand whether a policy change is responding to typical usage or to a relatively small group of extremely heavy workloads.
The same principle applies to token consumption. A small percentage of users may account for a disproportionate share of inference demand, particularly through long contexts, reasoning models, large files, agentic workflows, or repeated tool calls.
Again, I am not asking OpenAI to reveal confidential infrastructure details. Normalised or relative figures would be sufficient. For example:
A 128K-context interaction consumes approximately X times the compute of a typical 8K interaction.
or:
The heaviest 1% of conversations account for approximately X% of total inference workload.
That kind of information would make changes to usage limits much easier to understand.
At present, when limits change, users usually see the policy itself but have very little visibility into the workload distribution that motivated it. This can make an engineering or capacity-management decision appear arbitrary even when there may be perfectly reasonable technical reasons behind it.
If ChatGPT is becoming a mature computing service used for everything from occasional questions to hours-long research, coding, and continuous discussions, I think greater transparency about how people actually use the system would be valuable.
In short:
When usage limits change, could OpenAI show us not only the new limit, but also the workload model that made the change necessary?