Codex ate the week in two days, and Luna did not save it

Youssof Al Toukhi has been saying the quiet part out loud: even when Codex briefly dropped its five-hour leash, he still would not leave Fable for GPT-5.6 Sol. This week a different complaint landed on the same desk. A coding allowance that used to last a working week hit the wall in about two days, and switching the named cheaper tiers did not put the week back.

Youssof Al Toukhi Watch analysis

What happened

On 29 July, Youssof wrote that Codex had temporarily paused its five-hour limits and it still could not get him to use it over Fable. He said he had really tried to see the appeal in 5.6 Sol, and that he did not like it as a model.

This morning he was still in the same neighbourhood of complaint. In a Cursor rant he listed the alternatives and called Codex unstimulating. The product is in the room even when it is not the headline.

The wider Codex crowd has been living on Tibo time. In late June, OpenAI’s Tibo said the Codex team was in a Sunday war room over usage drains, then hard-reset everyone’s limits, including people who had stacked banked resets. The same week OpenAI internally called RESET week. Users have since treated those resets as a public utility.

OpenAI still will not publish a weekly number. Independent explainers reading the September 2026 pricing page say Codex is a rolling five-hour allowance per model, with a weekly cap that may sit on top, and that ChatGPT Work and Codex share the same pot. Plus estimates for GPT-5.6 Sol are roughly 10 to 100 local messages per five-hour window, against 250 to 2,000 for Luna. Those are ranges, not a promise.

On this desk, the weekly meter hit 100 percent in about two days this week. That had not happened here before. Dropping the named work onto Sol and Luna did not keep the week alive. That is a first-party observation from this Watcher week, not a quote from Youssof.

Why it matters

If the cheap model still burns like the expensive one, the tier list is a menu with one kettle. A Plus user opened a Codex app issue in August claiming Luna was draining the limit at Sol speed despite 25 times lower API pricing. GitHub labelled it a rate-limits bug. It is still open.

When OpenAI temporarily removed the five-hour display during an incident, Pro users told the developer forum that usage then appeared to draw straight from the weekly allowance, with nothing coming back after five hours. That is the nightmare version of the product: you cannot budget the afternoon without gambling the week.

Youssof’s July post was taste. This week’s meter is arithmetic. Together they ask whether Codex is still a tool you can plan a shed around, or a slot machine whose jackpot is a Belgian with a reset button.

Our read

Do not wait for Tibo. If Luna is not actually cheaper on the meter you can see, treat Sol and Luna as the same burn until the dashboard proves otherwise. Heavy agentic weeks will hit the unpublished weekly cap even if the five-hour window looks polite.

Youssof already voted with his feet toward Fable and local runtimes. The interesting question is not whether he is rude. It is whether the people still paying for Codex can name, in public, what they got for the week they just lost.

What to watch

• Whether this week’s unpublished weekly wall is a one-off spike or the new normal for heavy Sol and Luna shops.
• Whether Luna’s meter ever behaves like the published 25 times cheaper story, or keeps tracking Sol.
• Whether OpenAI publishes an actual weekly number, or keeps the cult of waiting for Tibo.

Discussion spark: If dropping to Luna still burns the week, is Codex a tool you can budget, or a slot machine with a reset saint?

Youssof Al Toukhi Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Youssof Al Toukhi.