Why Shared ChatGPT and Codex Limits Could Disrupt Research, Code Review, and Continuous Development
I’d like to share my concerns about the upcoming shared usage limits for ChatGPT and Codex, particularly regarding GPT-6 High in regular ChatGPT conversations.
I want to make something clear from the beginning: I’m not against integrating Work into ChatGPT. In fact, I think it could be an excellent improvement. My concern is about potentially losing independent access to high-quality reasoning in regular conversations.
1. My workflow: ChatGPT High is more important than the agent itself
I’m an independent developer working on a fairly niche and technically unconventional project. Because of its nature, my work requires extensive research, architectural discussions, hypothesis testing, and exploring approaches that often don’t have established solutions.
Approximately 80% of my work happens in regular GPT-6 High conversations, not in Codex or Work.
I use ChatGPT High to:
-
Research technical problems and explore possible architectural solutions.
-
Form, challenge, and refine hypotheses through long discussions.
-
Review source code and documentation using the GitHub connector.
-
Prepare and refine tasks before handing them to Codex.
-
Most importantly, review and verify the code produced by Codex, identify potential mistakes, and determine whether the implementation actually matches the intended architecture.
That last point is especially important.
For a technically unconventional project, implementing something is often easier than determining whether the implementation is actually correct. Agents can produce seemingly reasonable solutions that don’t necessarily match the real behavior of the systems being studied.
In many cases, the analysis and verification process is more important than the actual code generation.
This is why ChatGPT High is not simply a supplementary tool for me. It is the foundation of my development workflow.
And the ability to use it extensively, without competing with Codex’s usage limits, was one of the main reasons I chose ChatGPT as my primary AI tool.
2. Shared limits could eliminate the ability to continue working after Codex reaches its limit
Currently, when I reach my five-hour or weekly Codex limit, my work doesn’t stop.
I can continue using regular ChatGPT High to analyze existing implementations, review documentation, investigate technical problems, and prepare the next tasks for Codex.
Once the Codex limit resets, I already have a new set of well-researched and carefully planned tasks ready for implementation.
This separation is an extremely important advantage of the current system.
If regular GPT-6 High starts consuming the same shared budget as Codex, exhausting the agent’s allowance could also prevent me from continuing my research and planning.
And this isn’t necessarily solved by simply increasing the total shared allowance.
The problem is that two previously independent workflows would become dependent on the same limited resource.
Instead of continuing productive work, I could be forced to stop both implementation and analysis until the shared allowance resets.
Even before reaching that limit, I would have to constantly decide whether a discussion is worth spending tokens that might be needed for Codex later.
That turns a natural research and development process into continuous usage planning.
3. Regular conversations cannot simply be replaced with agents
Another important distinction is that regular ChatGPT High and agent-based workflows serve different purposes.
In my experience, agents such as Work and Codex are primarily oriented toward completing a defined task.
That is extremely useful when the objective is already understood.
However, when exploring an unfamiliar technical system, the objective itself may be unclear. Sometimes I need to spend hours investigating possibilities, questioning assumptions, discussing unusual approaches, or even deliberately exploring ideas that may turn out to be wrong.
Agents tend to approach these situations differently, often focusing on finding an actionable solution rather than maintaining a long, open-ended technical discussion.
This isn’t necessarily a flaw in agents. They are simply better suited to a different type of work.
I don’t want an agent to replace my technical discussions. I want to use agents after those discussions have produced a sufficiently well-defined task.
4. My experience with Claude shows why shared limits concern me
I’ve also experimented with Claude, which uses shared usage limits.
I don’t want to turn this into a comparison of which AI model is better. In fact, I currently prefer ChatGPT for many types of technical analysis, which is precisely why these changes concern me.
But I had one particularly frustrating experience with Claude.
I asked it to examine a few documents on GitHub and provide a relatively straightforward analysis.
That single request, involving connector usage and a fairly limited amount of research, consumed approximately 11% of my five-hour usage allowance.
For my workflow, where a single architectural discussion may involve reviewing many files, conducting numerous searches, and evaluating different hypotheses, that kind of unpredictable usage makes continuous work extremely difficult.
I’m not claiming that OpenAI will implement exactly the same accounting system.
However, this experience demonstrates why I consider independent access to regular high-reasoning conversations such a significant advantage.
If ChatGPT adopts similar limitations, one of the main reasons I prefer it over competing services could disappear.
5. Automatic tool usage is a strength of ChatGPT, not something I want to restrict
One of the things I particularly appreciate about ChatGPT is how naturally it can incorporate external tools into regular discussions.
During technical research, ChatGPT may perform numerous web searches, inspect GitHub repositories, retrieve documentation, and analyze considerable amounts of information.
This happens as part of the conversation, without requiring me to organize every individual operation manually.
And that’s exactly how I want it to work.
I have no complaints about the fact that ChatGPT autonomously decides when to use these tools. On the contrary, I consider this one of its strengths.
My concern is that, under a shared token-based allowance, this same behavior could make ordinary research conversations unexpectedly expensive in terms of usage limits.
Depending on the complexity of a question, the model might need a few searches or dozens of them.
I don’t necessarily know how many operations will be required, nor should I have to.
I don’t want to discourage the model from using tools, manually restrict its research, or constantly calculate whether another search is worth the remaining allowance.
Doing so would undermine the very reason I use ChatGPT for research.
The problem is not a lack of control over tool calls. The problem is the possibility that normal, productive use of these tools could become something I have to constantly budget and restrict.
This would fundamentally change the experience that currently makes ChatGPT so useful to me.
6. Intelligent UI introduces another uncertainty
I also want to mention the recent introduction of interactive UI elements in ChatGPT.
Visually, I think these features are impressive, and I understand why many users would appreciate interactive charts, diagrams, tables, and other components.
Personally, however, I rarely need them for my regular discussions.
Since this update, I’ve noticed that my long conversations seem to reach the hard conversation-length limit noticeably sooner. I’m referring specifically to the message telling me that the conversation has reached its maximum length and that I need to start a new chat.
This is only my personal observation, not a confirmed technical explanation. I cannot establish whether the interactive components are responsible.
However, it raises another concern about the future shared limits.
Interactive responses may require additional processing or generated content. If that additional work counts toward the new allowance, users could potentially consume more of their limited budget simply because ChatGPT chooses a more elaborate presentation format.
I don’t know whether this is how OpenAI intends to account for such content.
But it illustrates the broader issue: a regular conversation already involves many decisions made automatically by the model.
That automation is generally useful and welcome.
What concerns me is the possibility of introducing restrictive usage limits around a system that was previously convenient precisely because I didn’t have to think about those decisions.
7. Integrating Work into regular ChatGPT could actually be an excellent improvement
Despite these concerns, I genuinely support the idea of integrating Work more closely with regular ChatGPT conversations.
The current transition between Chat and Work isn’t particularly convenient.
When I need to preserve the context of a long discussion, I often have to create a separate branch in Work mode. This can become slow or unreliable with large conversation histories, and returning to the original style of discussion is not always straightforward.
A much better experience would be:
-
Discuss and research a problem in regular GPT-6 High.
-
Temporarily invoke Work for one or several specific tasks, using the existing conversation context.
-
Receive the results and immediately continue the same conversation in regular High, without creating branches or changing the entire workflow.
That would be a genuinely valuable improvement.
For example, I could spend time discussing an architectural change and then ask Work to perform a large GitHub documentation update that would be inconvenient to execute through ordinary connector calls.
Afterward, I could immediately return to regular technical discussions.
But integrating these capabilities into one interface does not inherently require merging their usage allowances.
I would strongly support the interface improvement while still opposing the loss of independent High usage.
8. This matters beyond programming
Although my main example involves software development, I don’t think this issue is limited to developers.
Regular ChatGPT conversations are useful for research, learning, decision-making, investigating complex questions, and countless everyday situations.
I have recommended ChatGPT to many friends precisely because of the quality and accessibility of its regular conversational experience.
Many of those people have absolutely no need for Codex or Work.
And while Instant may be sufficient for simpler tasks, my experience is that even relatively small research questions can benefit substantially from High reasoning.
For users who rely on those capabilities, replacing broadly accessible High with a restricted shared budget could be a significant downgrade, regardless of how many new agent features are introduced.
It would also make me reconsider recommending ChatGPT to others for the same reasons I have recommended it in the past.
9. What actually concerns me about this announcement
I understand that the exact implementation of the new limits has not yet been fully explained.
And perhaps the final system will be much more generous than I currently fear.
But that uncertainty is precisely what makes this announcement so frustrating.
I have spent considerable time building a workflow around ChatGPT because its current capabilities suit my needs extremely well.
I am satisfied with the quality of GPT-6 High, its ability to conduct extended technical discussions, its integration with external tools, and the separation between regular conversations and agent-based work.
I am not looking for a different way to use ChatGPT. The way I use it now already works extremely well.
That’s why the possibility of losing this workflow is so disappointing.
I’m not asking for detailed accounting of every tool call, nor am I asking for more ways to manually manage token consumption.
I don’t want to turn ordinary conversations into a resource-management exercise.
My concern is much simpler: I may lose the ability to work naturally and continuously with a model that currently meets my needs.
The announcement emphasizes that most users should not need to wait for their limits to reset more frequently than they do now.
But this doesn’t fully address my concern.
Even a generous shared allowance would not necessarily replace the advantage of having two independent ways to continue working.
Today, reaching the Codex limit does not mean I have to stop researching, analyzing, and planning in ChatGPT High.
Under the unfavorable interpretation of the announced changes, that independence could disappear.
I also find it difficult to understand why such a significant potential change to an established workflow is being announced without clearer information about what regular High users should actually expect.
I’m not claiming that the worst-case scenario will happen. I genuinely hope it won’t.
But the possibility alone is enough to make me reconsider how much I can rely on ChatGPT as my primary long-term working environment.
And I find that unfortunate, because I currently have very few complaints about the product itself.
Conclusion
For my particular workflow, regular GPT-6 High is not a secondary feature of ChatGPT. It is the primary reason I subscribe.
I spend far more time researching, discussing, planning, reviewing, and validating work than actually running agents.
The ability to continue those activities after exhausting Codex’s limits is essential.
I genuinely appreciate ChatGPT, and I currently consider it a better fit for my analytical workflow than the alternatives I’ve tried. I don’t want to switch services, and I’m not opposed to new features.
I support seamless Work integration into regular conversations. I do not support making regular High reasoning dependent on the same exhausted budget as Codex.
What I hope OpenAI understands is that many users don’t see regular ChatGPT as a lightweight preliminary step before agent execution.
For some of us, the conversation itself is the most valuable part of the product.
I would strongly prefer to see that experience preserved, with agent capabilities added as a complementary feature rather than becoming a competing use of the same limited resource.
The concern isn’t simply that the new system might provide fewer tokens. It’s that it could take away the freedom to continue meaningful work in the way that already works best for us.