A ChatGPT Plus subscriber began a GPT-6 Astra session with a full allowance at 2:51 p.m. Twenty minutes later, they say, it was gone.
In a September 6 report on OpenAI’s Codex GitHub repository, the user described reaching the limit during what they said were ordinary, small-scale tasks. “A 20-minute window is simply too short to accomplish meaningful work,” they wrote.
Other Astra users describe an even shorter encounter: one request, an exhausted allowance and work still unfinished. Complaints across Reddit, X and OpenAI’s developer forum are raising an awkward question for people paying to hand tasks over to AI: how much can it actually finish before access runs out?
The complaints centre on Astra in Codex and ChatGPT Work, where tasks draw on a shared usage allowance. OpenAI has already announced improvements to consumption for some Astra workloads. The accounts reviewed by Search Engine Watch are user reports, rather than independently reproduced measurements, and do not establish how widespread the problem is.
On September 7, X user Md Ismail Šojal summed up the frustration with a post captioned: “The most expensive ‘hello’ I’ve ever seen. GPT-6 Astra on Codex.” Original post.
In a Reddit thread about Astra on Plus, user HenryofSAC said they had used all their saved usage resets in a day, having previously never needed one. A single prompt could exhaust their allowance, they said.
Another commenter, Pitiful-Hearing-5352, reported using the entire five-hour allowance in three minutes after asking Astra to review a project.
“And the sad thing is that it’s not done,” they wrote.
That detail gives the complaints their force. An expensive task can still be worthwhile if it produces useful work. An exhausted allowance leaves much less to show for itself when the requested result has not arrived.
A separate September 5 Codex bug report raises a more specific concern about purchased credits. Its author says their balance kept falling while the application repeatedly tried to restore a failed connection.
The reporter describes roughly three credits disappearing every ten seconds during a reconnection loop that could last about ten minutes. They say no useful output appeared and no files were changed during that period. Once Codex switched to an alternative connection method, the task began progressing normally.
The author says they monitored the purchased-credit balance on OpenAI’s usage page after exhausting their included allowance. That makes this a different allegation from simply finding Astra expensive: they describe additional credits being consumed during an apparent stall.
Search Engine Watch has not independently reproduced the behaviour or established whether the balance changes represented fresh charges, delayed reporting or work continuing outside the visible interface. The report warrants investigation; it does not prove a general billing fault.
OpenAI’s published limits help explain why some subscribers encounter a much shorter session than they expected. Its estimates put Astra at 5–45 local messages per five-hour period on Plus, compared with 10–100 for GPT-5.6 Sol. Those figures vary by workload and are not guaranteed message allowances.
The five-hour window is the period over which usage is measured, rather than a promise of five hours of continuous work. Prompt length alone is a poor guide: context, reasoning, tools, retrieval and caching all affect consumption. Work and Codex also share usage.
OpenAI says an active turn can continue after a limit is reached, subject to fair-use restrictions. Users can buy additional credits when their included allowance is exhausted. Those provisions make it especially useful to establish exactly where the reported unfinished tasks stopped.
OpenAI’s Tibo announced Astra usage improvements on September 6, saying some power-user workloads could draw three to four times less from a subscription without a quality change. He described the benefit as applying to the “long tail” of usage, without specifying the affected workloads or an average reduction. The announcement did not promise a blanket increase in everybody’s allowance.
Experiences also differ. In the same Reddit discussion, user gregpeden said a two-hour website-building task on Astra’s medium setting consumed about 11% of their Pro 5x weekly allocation. That is a different subscription and workload from the Plus complaints, but it demonstrates why the fastest reported depletion should not be treated as a universal rate.
The dissatisfaction extends to what access itself is worth. In a September 7 post on OpenAI’s developer forum, a Plus subscriber argued that offering the newest model through heavily metered Work and Codex made that access less meaningful.
For marketers handing over a site audit, a research assignment or a development task, the practical question is how much finished work the subscription buys. A model can be impressive and still be difficult to budget for. The users complaining about unfinished jobs have already encountered that distinction.
Start the conversation by posting the first comment