When Every Message Counts: Is ChatGPT Changing the Way We Think?

Oct 2

3 min read

I started a conversation with ChatGPT this week about a fairly simple problem: why were some of my scheduled tasks becoming disabled?

Several messages later, we had identified a much bigger issue and redesigned the underlying automation architecture.

This is exactly how I use AI.

I don't always know the right question at the beginning. I ask something, challenge the answer, add context, change direction and gradually expose the real problem.

In other words, I "CHAT" with ChatGPT.

And increasingly, that has a cost.

It isn't just a Pro issue

OpenAI now applies different usage allowances across plans, models and features.

The exact mechanics vary. Free and Go users currently have unlimited everyday text conversations using Luna, subject to safeguards, while tools have separate limits. Plus users have limits around higher reasoning and agentic capabilities. Pro and Business plans have explicit allowances for premium models. Work and Codex are accounted for differently again.

One particularly visible example arrives on 30 October 2026, when affected Pro 200 customers (me) move from 200 to 100 GPT-6 Pro messages per week.

But the important issue is not the number 100.

It is the behaviour that message limits encourage.

The rational response is to stop chatting

If ten short exchanges consume ten messages while one carefully constructed 2,000 word prompt consumes one, the economically rational behaviour is obvious.

Don't explore the problem conversationally. Prepare it first. Provide all the context. Anticipate the questions. Ask for everything in one go.

That is efficient.

It is also remarkably similar to writing a specification for a machine.

And that matters because prompt efficiency and thinking efficiency are not the same thing.

The first question is rarely the real question

In consulting, coaching, leadership and strategy, good conversations evolve.

An operational problem becomes a governance problem. A technology question exposes a process failure. A challenge to one assumption opens an entirely different option.

You cannot always write the perfect opening prompt because you don't yet know what you need to know.

The conversation discovers it.

In my case, I could have saved messages by starting with:

Investigate Scheduled Tasks, research OpenAI's latest architecture, diagnose the failure, assess event driven alternatives and redesign the PROBOS automation operating model.

Excellent prompt.

There was just one problem.

I didn't know that was the question when I started.

Cost per message may be the wrong measure

There is an entirely reasonable economic argument for usage limits. Frontier AI requires expensive computing resources, and more capable models should increasingly accomplish more in each interaction. But there is a difference between using AI as an answer engine and using it as a thinking partner.

For an answer engine:

Know what you want. Ask efficiently. Get the answer.

For a thinking partner:

Explore. Challenge. Reframe. Discover.

Message based limits naturally encourage the first behaviour.

Individual users across OpenAI's own community and Reddit are already discussing limits, monitoring usage and changing how they select models and structure conversations. That is anecdotal evidence, not proof of a widespread behavioural shift, but the incentive itself is straightforward.

Perhaps AI literacy is changing

We may now need another skill beyond prompt engineering.

Conversation engineering.

Knowing when to use a lower cost model. Knowing when deeper reasoning is worth the allowance. Knowing when to consolidate ten exploratory questions into one serious piece of analysis. And knowing when not to optimise, because another conversational turn might reveal that we are solving the wrong problem.

That is probably sensible.

But there is an irony.

The better AI becomes as a thinking partner, the more its usage economics can encourage us to communicate with it like an old fashioned computer: prepare the specification, minimise interactions, maximise output.

At PROBOS, I think the more useful question is therefore not:

How do we get the most output from every AI message?

It is:

How do we design the interaction so humans and AI reach better decisions together?

Those are not the same thing.

Limits and model availability change frequently. Product details in this article were checked against OpenAI's published information on 1 October 2026.