AI-assisted development
Not Making Money With AI Yet? Get More From Your Codex Subscription
Two bookmarks can help you time ambitious Codex work. The real return still comes from what you ship after the reset.
If AI is not earning or saving you at least the price of your Codex subscription, the answer is not to squeeze out more tokens for their own sake. The answer is to turn the allowance you already have into one shipped result, one useful automation, or one piece of evidence that moves you closer to a customer.
Usage still matters. A capable agent is easiest to underestimate when it sits idle, and easiest to waste when every session begins with a vague prompt. Codex limits make that waste visible. Public resets make the timing stranger: sometimes fresh capacity arrives before the normal window, sometimes it can be saved for later, and sometimes a hint is only a hint.
I keep the decision simple. I check my own usage first. Then I use Codex Forecast to see whether there is a live reset signal. If there is, I open the original post on Tibo Sottiaux’s feed and read the wording for myself.
Start with your own meter
Your account is the source of truth. OpenAI’s current Codex pricing and usage guidance explains that consumption changes with the model, task complexity, context, reasoning, tools, caching, and whether work runs locally or in the cloud. Two prompts that look similar can use very different amounts of the allowance.
Check the usage dashboard before planning a large session. In the CLI, use /status. That tells you what is actually available to your account; a public announcement cannot do that.
This also keeps the model choice honest. Hard, ambiguous product work may justify the strongest model. Routine transformations, extraction, and high-volume cleanup usually do not. OpenAI’s own guidance recommends tighter scope, fewer irrelevant sources, fewer unnecessary MCP servers, and a smaller model when it can still produce the result you need.
The goal is not to spend every unit of capacity. It is to spend the right capacity on work that can change something.
Use Codex Forecast as a clue, not a countdown
Codex Forecast is an independent LunarWerx project that keeps a public record of Codex resets and estimates the chance of another one over the next six and twenty-four hours. It cross-checks public announcements, a monitored account, a reset tracker, and OpenAI’s status history. It also links every recorded event back to its source.
At the time I wrote this, the page listed 44 reset events, eight during the previous 30 days, and an average recent wait of 3.8 days. Those are dated observations, not a schedule. The page is explicit that it is not affiliated with OpenAI, has no inside information, and cannot turn an irregular sequence into a reliable countdown.
That honesty is the useful part. Its own backtest says the broad twenty-four-hour probability has performed about as well as repeating the historical average. The shorter window and the live-signal card are more actionable because they incorporate Pacific working hours and recent language from Tibo’s feed.
Read the forecast without fooling yourself
- If the only argument is “the average says we are due,” do not wait. An average is not an appointment.
- If a live card links to a post, check the post’s timestamp and timezone.
- Look for whether the post describes a hint, an incoming reset, a completed reset, or a banked reset.
- Before starting a large task, return to your own usage dashboard and confirm the capacity is there.
Follow Tibo for the signal, not the entertainment
Tibo leads Codex and uses his public feed for the kind of operational information that does not always fit neatly into product documentation: reset announcements, banked resets, rate-limit investigations, reliability updates, model releases, and examples of what people are building.
The difference between those posts matters. On August 22, he announced that a banked reset had landed for paid users. A banked reset is capacity you can keep until you need it; it is not the same as an immediate refill you must use now. The next day, he described several rate-limit inefficiencies involving image-heavy long sessions, Computer History, and conversation-title generation, alongside another reset plan.
That second kind of post is as valuable as the refill. It can explain why a workflow is consuming more than expected and keep you from “optimizing” the wrong part of your process.
The feed is also a useful product notebook. Tibo has shown how a clarification-first instruction can become a reusable Codex skill, and he has highlighted builders who turn a narrow distribution problem into a small hosted product. Neither example guarantees revenue. Both reinforce the same pattern: find concrete friction, make the useful behavior repeatable, and ship it.
I would not monitor the feed all day. Let the forecast or an alert tell you that something relevant happened, then open the original post. Read the verbs. “Coming,” “landed,” and “banked” imply different decisions.
Keep a capacity backlog
A reset is only leverage when useful work is already waiting. If fresh capacity arrives and you spend the first hour deciding what to ask, the timing advantage is gone.
Keep a short list of assignments that are valuable, bounded, and easy to verify. These should not be random chores invented to consume a quota. They should be work you intend to finish whether or not an early reset appears.
Good work for a high-capacity window
- Complete one sellable workflow end to end, including the failure state and deployment.
- Turn a repeated manual task into a small tool or automation you can measure.
- Run the real product flow in a browser and fix the highest-impact usability problems.
- Add the tests, onboarding, and operating notes that make a prototype usable by someone else.
- Ask a fresh session to review the plan or implementation and return evidence, not reassurance.
Write each item as an outcome with boundaries and evidence. “Improve the app” is not ready. “Make checkout recover from a declined card without losing the cart, preserve the current payment provider, and verify the flow in a local browser” is ready.
This is how more usage becomes useful. You are not giving Codex a larger pile of work. You are giving it a prepared sequence of complete outcomes.
Use a ten-minute weekly ritual
Once or twice a week, check the dashboard and decide whether your current pace matches the work you want to finish. If you have plenty of capacity, start the most valuable prepared assignment. If you are close to a limit, check the forecast’s short window and any live signal, then verify the linked Tibo post.
If a reset is explicitly incoming, use the remaining allowance on a smaller backlog item while you wait. If a reset has landed, confirm it in your account and start the larger task. If nothing concrete is happening, keep working at the pace your meter supports. Do not reorganize your day around a probability band.
At the end of the session, record what changed: the URL you shipped, the test that now passes, the customer workflow you completed, or the hours the automation will save. That record is more useful than a screenshot of a replenished quota.
A reset is not the return
If AI is not making money for you yet, use Codex to create evidence before you use it to create more software. Put a working tool in front of one person. Automate a task you already perform. Publish a small product. Ask for payment. Measure whether anyone returns.
Two bookmarks can help you avoid leaving useful capacity untouched: the independent forecast for timing and Tibo’s feed for the original signal. Neither one can make the subscription worthwhile for you.
The return comes from the thing you ship, the time you save, and the decision you can now make with better evidence. If those results are not appearing after a few weeks, change the work or change the plan. Chasing another reset will not solve it.