I would like to propose two changes to Codex subscriptions: a higher-usage Pro tier at approximately $400 per month with 50× usage, and a dedicated monthly allowance for a long-running autonomous research session.
The first would address an immediate capacity gap. The second would explore a subscription model better suited to tasks that require days of independent work rather than frequent human interaction.
Why the current usage limits leave a gap
The current 20× Pro plan can be quite generous for interactive coding, especially when users optimize their workflows. For example, a more capable main agent can focus on difficult reasoning and research decisions, while less expensive subagents handle routine implementation, verification, experiments, and literature searches.
With effective task delegation and careful token management, the weekly allowance can comfortably support our ordinary interactive workload, sometimes with capacity left over.
However, continuous autonomous research has a different usage pattern. A task such as investigating an open mathematical problem may require the model to work independently for several days, developing ideas, testing approaches, checking intermediate results, and revising its strategy. Even an efficient workflow consumes substantial resources when it runs around the clock.
The issue is therefore not simply whether the current plan provides enough usage for heavy interactive coding. It is whether the subscription can support sustained autonomous work.
What we are seeing in our own research workflow
We have been developing a local Codex workflow for automated research on open mathematical problems, with an emphasis on both research performance and cost efficiency. We currently use GPT-6-Astra Max as the main agent and GPT-5.6-Luna Max for most subagent tasks.
The main agent focuses on generating ideas, choosing research directions, and integrating results. Subagents handle work such as verification, experiments, and literature review whenever those tasks do not require the main model’s full capabilities. We have also worked extensively on preserving task continuity, reducing unnecessary token consumption, and making subagent work more efficient and reliable.
In our testing, this setup has maintained continuity and continued making progress through repeated context compactions. We believe it has also produced significant progress on an important open mathematical problem, and we are working toward a complete, verified solution.
Despite these optimizations, the workflow still consumes approximately 1% of the 20× plan’s weekly usage allowance per hour. At that rate, a full weekly allowance supports about 100 hours—just over four days—of continuous research, rather than an entire week.
For this workload, the immediate bottleneck is how long we can keep the research process running within the subscription’s limits. Improving efficiency helps, but it does not eliminate the capacity gap.
Proposal: A ~$400/month plan with 50× usage
A Pro tier offering approximately 50× usage for around $400 per month would help address this gap.
Assuming the quota scales proportionally, a 50× plan would provide 2.5 times the allowance of a 20× plan. At our current usage rate, that would be enough to support continuous research throughout the week, with some headroom for more demanding phases of the work. Any shorter-term usage limits would also need to accommodate sustained operation.
I believe this tier could appeal to both users developing autonomous research workflows and heavy Codex users who need more capacity for demanding coding projects.
The existing tiers would still have a role. A 5× plan may be sufficient for lighter workloads, while 20× can serve substantial interactive use. A 50× option would address users whose workloads go beyond those patterns, without requiring one subscription tier to accommodate every use case.
As models become more useful for independent, multi-day work, there is value in offering subscription options that support that pattern explicitly.
Additional proposal: A monthly autonomous-research allowance
A second, more experimental idea is to give users on the highest subscription tier a monthly allowance for one designated autonomous task session that does not count against their standard usage quota.
The basic rules could work as follows:
-
The user designates the session before it starts. The monthly allowance is reserved for that session, preventing it from being assigned to multiple sessions simultaneously.
-
The session runs without additional user instructions. The user defines the task at the beginning but cannot add new requests or redirect the session while it is running under this allowance. The user can still stop it at any time.
-
Completion or interruption consumes the allowance. Once the task finishes—or the user manually interrupts it—the allowance is considered used. Stopping a task would not restore the allowance for another attempt.
-
A monthly reset does not automatically terminate an unfinished task. The user can apply the new month’s allowance to the existing session or allow it to continue against the standard usage quota. Continuation would remain subject to whichever resource limits apply.
This would be an exemption from the standard quota, not necessarily a promise of unlimited compute. A separate, clearly disclosed resource budget and limits on concurrent work could help control costs and prevent abuse. The goal would be to make sustained autonomous work practical, not to let ordinary interactive usage be repackaged as a never-ending research session.
I recognize that this idea needs refinement, particularly around resource limits, task boundaries, subagent usage, and sessions that span multiple months. Because of the potential cost, it may only be appropriate for the highest subscription tier.
Nevertheless, I think it is worth exploring. A dedicated allowance could let users initiate a substantial research task with greater confidence that it can continue without exhausting the quota they also need for everyday work.
Closing thought
These proposals address the same underlying need in different ways: a higher-usage plan would provide more capacity immediately, while a dedicated research allowance would offer a different way to support long-running tasks.
The broader request is to make Codex subscriptions suitable not only for human-supervised coding, but also for sustained autonomous research. For users building these workflows, the ability to keep a capable system working continuously can be just as important as the capability of the model itself.
