Users report Astra and Codex caps halt long-running agent work

• Astra’s subscription caps can interrupt extended Codex work, as vikhyatk's usage report and altryne's overnight run show.
• A 30-minute Astra build of a voxel scene logged roughly 6 million tokens and nearly exhausted its five-hour limit, according to cedric_chee's voxel report.
• An update can draw up to 3-4x less subscription usage for some power users on the "long tail," per thsottiaux's update.
• Banked resets are one-time quota refreshes rather than permanent capacity, according to OpenAI's reset guide and thsottiaux's reset reply.

OpenAI's Astra announcement places the model across paid ChatGPT plans, Codex, the API, Azure, and Bedrock. The reset guide says a full reset also moves the weekly reset date. In a 64-subagent experiment, a 19-minute run consumed 5% of a Pro 5x weekly allowance, daniel_mac8's follow-up reported.

Capacity becomes an availability problem when an agent is meant to outlast its operator. altryne said in altryne's post that an Astra task stopped after two hours, despite their expectation that it would continue overnight.

koltregaskes said one day on Astra Light nearly exhausted the weekly allowance on a Pro 20x plan. cedric_chee separately reported repeated limit hits 10 minutes apart while doing no complex work.

An update from thsottiaux said long-tail power-user sessions could draw up to 3-4x less from a subscription without a quality change. It did not define the session length or context threshold covered by “long tail.”

In thsottiaux's effort guidance, thsottiaux said Astra Low outperforms Sol High and suggested moving Sol High users down to Low or Medium. They also put Astra at about 50% cheaper per task at roughly equivalent output in thsottiaux's cost reply. One reply from lateinteraction asked whether that calculation holds for workflows whose median context exceeds 100,000 tokens lateinteraction's question.

A reported Codex configuration raises max_concurrent_threads_per_session from a default of three to 64. daniel_mac8 said in daniel_mac8's demonstration that OpenAI tested Astra with up to 64 subagents.

daniel_mac8 put a 19-minute run at 5% of a Pro 5x weekly limit in daniel_mac8's follow-up, and said the Codex UI became laggy during the experiment.

OpenAI's banked-reset guide calls a reset a saved, one-time usage-limit reset, not purchased credit or a permanent increase to a plan’s allowance. A full reset refreshes both the five-hour and weekly windows, then changes the weekly reset date.

thsottiaux said in thsottiaux's reply that usage is no different before and after a reset.

Codex’s experimental context-management mode saves notes across context windows and searches earlier messages and tool outputs, according to reach_vb's configuration note. It is off by default; the note says it requires an updated Codex client, Astra, a Plus, Pro, or Pro Lite sign-in, and a new task.

A post from aibuilderclub_ described lower quota draw as user reports, not a guarantee, in aibuilderclub_'s caveat.