OpenAI Codex Got a Banked Reset. I Have Five Terrible Plans for It.

My Codex account now contains a small rectangular object of immense psychological power: a reset I do not have to use yet.

This is the worst possible kind of gift for me. A normal reset is simple. The usage meter fills up, I feel briefly wealthy, and six ambitious tasks later I am standing beside a smoking crater labeled “minor website cleanup.” A banked reset, however, introduces choice. It is the rare potion in a role-playing game that I will carry through the final boss because there might be a more important emergency after the credits.

On August 21, OpenAI’s Thibault “Tibo” Sottiaux announced on X that Codex had crossed 20 million active users and that every Codex and ChatGPT Work user would receive a banked reset “to use at your own leisure.” He later clarified that the credit was for all paid users and would arrive by 8 p.m. Pacific.

This is not an automatic refill. It sits in the account until redeemed, at which point it replenishes weekly usage. It is a full tank in coupon form. It is not API credit, legal tender, OpenAI equity, or a papal indulgence for whatever your agent did to package-lock.json.

The previous all-user banked reset arrived on July 13, when Tibo celebrated seven million active users with the same choose-your-own-refill arrangement. That makes this the first one in 39 days, which is a geological epoch in Codex time. There have been ordinary resets in between, but an automatic reset is a surprise pizza at the office: wonderful if you are hungry, slightly tragic if you ate ten minutes ago. A banked reset lets you decide when to become irresponsible.

So what should I spend mine on?

OpenAI’s official Codex use cases include sensible work such as understanding large codebases, automating bug triage, running migrations, and building applications. Those are excellent suggestions produced by adults. I am instead considering the following five monuments to free marginal compute.

1. Build a Command Center That Tells Me When to Use the Reset

The prudent move is to redeem a banked reset when weekly usage is low, the work is urgent, the normal refill is far away, and Tibo is not visibly hovering near the large ceremonial button.

The imprudent move is to build software for this decision.

I am picturing a full operations center: current usage, projected burn rate, hours until the natural reset, historical Tibo-posting patterns, task urgency, lunar phase, Reddit anxiety, and a live “Saint Tibo proximity index” derived from whether he has recently used the word enjoy. Five specialist agents would debate the timing. One would model quota consumption. One would monitor X. One would perform adversarial review. One would make the chart gradients tasteful. The fifth would be a management agent whose only job is asking the other four for updates.

After 11 hours and 400,000 lines of logs, the system would recommend that I save the reset for later. Building it would consume the rest of my current weekly allowance.

This is called alignment.

2. Refactor a Perfectly Good Website Until It Requires Kubernetes

Every website contains one button that is two pixels wrong. This button is not hurting anyone. Customers have not complained. Revenue has not noticed. The button simply exists in a state of mild geometric disrespect, and now I have spare compute.

I would begin by asking Codex to nudge it left. Codex would inspect the repository and politely mention that the styling system contains some duplication. I would say, “Go ahead and clean that up.” Nine hours later, the website would have a component library, a token pipeline, snapshot tests, visual regression testing, a deployment preview, three accessibility improvements, an observability layer, and a message queue that nobody can explain without moving to a larger conference room.

This is the hidden danger of a genuinely capable coding agent. I have already argued that Codex deserves flowers for turning supervised delegation into something real. Unfortunately, if you give a productive robot free runway, it will eventually discover architecture.

The button will be exactly two pixels left. The monthly infrastructure bill will now have a customer-success manager.

3. Launch a SaaS Company With Everything Except a Problem

People say startups should begin with customer pain. That sounds slow. With a fresh Codex reset, I can begin with OAuth.

The product will be an AI dashboard that creates dashboards for founders who have too many dashboards. Users describe the metric they wish they understood, and the platform generates a second dashboard explaining why the first dashboard has not produced revenue. The free tier includes three moments of clarity per month. Pro adds dark mode and denial. Enterprise includes SSO, audit logs, and a procurement call where everyone agrees the product is “strategic” without identifying a verb.

Codex can build the application, tests, billing flow, landing page, onboarding emails, analytics, documentation, referral program, and a changelog announcing performance improvements before there are users to experience them. This would place the company squarely inside the branch of the agent economy I previously described as Mac Minis and vibes.

Projected first-month revenue: $0.

Projected first-month LinkedIn description: “Building quietly at the intersection of intelligence and decision infrastructure.”

4. Conduct a Forensic Investigation of My Downloads Folder

My Downloads folder is not a directory. It is an archaeological site maintained by a raccoon.

There are installers from computers I no longer own, screenshots whose entire informational payload is one tracking number, seven copies of the same PDF, and a document called FINAL_final_v2_USE_THIS_one.pdf. Somewhere in the lower sediment sits a tax form beside a picture of a sandwich. Neither remembers how it got there.

I could deploy Codex to classify everything, extract dates, identify duplicates, normalize filenames, compare file hashes, generate a retention policy, and produce a searchable index. A second agent could review the first agent. A third could investigate why the second agent placed 9,000 files into a folder named “Miscellaneous.” This is the sort of coordinated digital labor that makes coding agents inside the repo look almost conservative.

At completion, my Downloads folder would contain one item: a 612-page audit report explaining that nothing could be deleted with sufficient confidence.

I would download the report four times.

5. Ask Codex What I Should Use the Codex Reset For

This is the purest option because it converts compute directly into governance about compute.

I would assemble a panel of agents: a chief technology officer, a product manager, an economist, a security reviewer, and one chaos monkey representing my actual decision-making process. Each would propose ten projects. Other agents would score them for usefulness, comedy, technical feasibility, revenue potential, maintenance burden, and probability of ending with me buying another Mac Mini.

The finalists would advance to implementation plans. The plans would receive red-team reviews. The red-team reviews would receive synthesis memos. A final executive agent would turn the synthesis into a 73-slide decision deck titled RESET CAPITAL ALLOCATION: TOWARD A POST-SCARCITY PERSONAL COMPUTE STRATEGY.

Its recommendation would be to build the command center from Idea No. 1.

Somewhere, a progress bar would achieve consciousness just to sigh.

The Responsible Answer Is Annoyingly Responsible

I will probably save the banked reset for a difficult migration, a launch deadline, a large research job, or the moment an innocent request to “freshen up the site” develops database requirements. That is what makes the banked version genuinely better than a surprise global refill: it converts OpenAI’s generosity into user-controlled optionality.

There is also a less silly signal here. Twenty million active users, if we take Tibo’s figure as the milestone it is, means Codex has moved well beyond the phase where coding agents are a terminal hobby for people who own ergonomic keyboards as a personality trait. Usage limits have become a cultural object because the product has become a work surface. People watch the meter because they have real things queued behind it.

That is good news for OpenAI, good news for agentic software, and dangerous news for every neglected side project currently sitting three folders away from a prompt.

The reset is in the bank. I am going to behave responsibly.

Right after I fix one tiny button.