ChatGPT Is Retiring Custom GPTs. Will the Replacement Apps Cost Creators More?

Our Second Experiment: A Full HTML Article

For the next test, we changed both the model and the workload.

We switched to GPT-6.1 Sol Medium, which OpenAI positions as a lower-cost option than Astra for complex professional work.

Then we asked the plugin to complete a task much closer to our normal publishing workflow.

The plugin completed the assignment.

Immediately afterward, ChatGPT told us:

The interface offered several next steps, including upgrading, purchasing additional credits, or waiting for the included allowance to reset.

We chose not to buy additional credits.

The Five-Hour Limit Doesn’t Mean Five Hours of Clock Time

This is one of the easiest parts of ChatGPT’s usage system to misunderstand.

OpenAI explains that Work and Codex allowances are not simply a fixed number of prompts, and a five-hour usage limit can be exhausted before five literal hours have passed.

Consumption varies based on factors such as the model, task, context, reasoning setting, input and output size, and the tools being used.

OpenAI’s own Plus estimates show how wide that variation can be. It currently estimates roughly 5–45 local messages with GPT-6 Astra in a five-hour window, compared with approximately 15–160 using GPT-6.1 Sol.

Those are estimates, not guaranteed prompt counts, and more involved Work tasks can consume more allowance than simple local messages.

That means a creator can’t look at “5% remaining” and reliably translate it into five more prompts, 30 more minutes, or one more article.

What Our Experiment Actually Proves

There is a temptation to look at these two tests and conclude that migrated Custom GPTs are going to be dramatically more expensive.

We don’t have enough evidence to make that claim.

We also can’t claim that our first 30-second script alone caused the entire observed 9% to 5% movement.

A proper cost comparison would require a controlled benchmark: the same source files, the same prompt, comparable models, equivalent reasoning settings, a fresh usage window, and repeated trials.

That’s a test worth doing.

Why Creators Should Care About the Difference Between Chat and Work

OpenAI now distinguishes between ordinary ChatGPT conversations and more agentic Work tasks.

Chat is intended for things like asking questions, brainstorming, searching, and conversational help. Work is designed for larger jobs such as research, analysis, and creating documents, spreadsheets, presentations, reports, and other deliverables.

That’s a meaningful distinction for creators because many Custom GPTs were built precisely to generate substantial deliverables.

A blogger may use one to turn research into a formatted WordPress article. A video creator might use another to create scripts, titles, descriptions, and social posts. A small business might use one to process reference files and create marketing material.

Those aren’t novelty prompts. They’re production systems.

If replacement workflows now operate inside a more constrained shared allowance, creators need to understand that before migrating the systems they depend on.

Our Second Experiment: A Full HTML Article

For the next test, we changed both the model and the workload.

We switched to GPT-6.1 Sol Medium, which OpenAI positions as a lower-cost option than Astra for complex professional work.

Then we asked the plugin to complete a task much closer to our normal publishing workflow.

The plugin completed the assignment.

Immediately afterward, ChatGPT told us:

The interface offered several next steps, including upgrading, purchasing additional credits, or waiting for the included allowance to reset.

We chose not to buy additional credits.

The Five-Hour Limit Doesn’t Mean Five Hours of Clock Time

This is one of the easiest parts of ChatGPT’s usage system to misunderstand.

OpenAI explains that Work and Codex allowances are not simply a fixed number of prompts, and a five-hour usage limit can be exhausted before five literal hours have passed.

Consumption varies based on factors such as the model, task, context, reasoning setting, input and output size, and the tools being used.

OpenAI’s own Plus estimates show how wide that variation can be. It currently estimates roughly 5–45 local messages with GPT-6 Astra in a five-hour window, compared with approximately 15–160 using GPT-6.1 Sol.

Those are estimates, not guaranteed prompt counts, and more involved Work tasks can consume more allowance than simple local messages.

That means a creator can’t look at “5% remaining” and reliably translate it into five more prompts, 30 more minutes, or one more article.

What Our Experiment Actually Proves

There is a temptation to look at these two tests and conclude that migrated Custom GPTs are going to be dramatically more expensive.

We don’t have enough evidence to make that claim.

We also can’t claim that our first 30-second script alone caused the entire observed 9% to 5% movement.

A proper cost comparison would require a controlled benchmark: the same source files, the same prompt, comparable models, equivalent reasoning settings, a fresh usage window, and repeated trials.

That’s a test worth doing.

Why Creators Should Care About the Difference Between Chat and Work

OpenAI now distinguishes between ordinary ChatGPT conversations and more agentic Work tasks.

Chat is intended for things like asking questions, brainstorming, searching, and conversational help. Work is designed for larger jobs such as research, analysis, and creating documents, spreadsheets, presentations, reports, and other deliverables.

That’s a meaningful distinction for creators because many Custom GPTs were built precisely to generate substantial deliverables.

A blogger may use one to turn research into a formatted WordPress article. A video creator might use another to create scripts, titles, descriptions, and social posts. A small business might use one to process reference files and create marketing material.

Those aren’t novelty prompts. They’re production systems.

If replacement workflows now operate inside a more constrained shared allowance, creators need to understand that before migrating the systems they depend on.

One Important Enterprise Detail

There is another reason creators should be careful when reading statements about plugin pricing.

OpenAI says that in Enterprise workspaces, using a migrated GPT without apps in Instant Chat has no incremental cost. It separately says that creating and editing plugins through a conversation uses Work or Codex, and that actions performed in Work consume credits.

That’s useful information, but it isn’t a blanket promise that every plugin interaction on every plan works the same way.

We’re testing on Plus, not Enterprise.

The difference between a lightweight plugin interaction in Chat and a plugin being used as part of a Work task needs to be easier for ordinary users to understand.

Five Questions We Want OpenAI to Answer

These aren’t accusations about OpenAI’s motives.

They’re practical questions from someone actually trying to migrate a production workflow.

What Creators Should Do Before Migrating Their GPTs

OpenAI itself recommends testing migrated plugins against familiar prompts. We’d go a little further.

What We’re Doing Next

For now, we’re holding off on migrating additional production systems.

That isn’t because we’ve decided plugins are bad.

Plugins may ultimately be much more capable than Custom GPTs. Combining reusable instructions, reference material, apps, and advanced tools could create workflows that weren’t practical before.

But capability and predictability are two different things.

If an independent creator is going to build their publishing system around these tools, they need to know roughly what a routine job will consume and when it may move from included subscription usage into additional credits.

Our two tests don’t settle that question.

They do give us a reason to keep testing it.

The Bottom Line

We’ll continue comparing plugin workloads, models, reasoning settings, and ordinary ChatGPT workflows as the Custom GPT retirement gets closer.