OpenAI resets GPT-6 Astra usage for paid subscribers

• OpenAI restored GPT-6 Astra capacity across paid subscriptions after thsottiaux's announcement promised a global reset, then thsottiaux's completion post said it had reached everyone.
• Future resets have no fixed timetable, according to thsottiaux's reply, which said, “There is no schedule, only resets.”
• Subscribers reported sharply different exhaustion patterns, with bridgemindai's usage screenshot showing a depleted weekly allowance alongside a fully available five-hour window.
• OpenAI said a long-tail optimization can draw up to 3-4x less subscription usage with no quality change in thsottiaux's update, while aibuilderclub_'s configuration post surfaced experimental context management for long runs.

OpenAI’s banked-reset guide defines a reset as a temporary, expiring recovery of capacity, separate from a permanently larger plan allowance. An OpenAI Developer Community report describes a newly restored five-hour limit falling to zero twice in minutes, while the experimental context-management setting offers a distinct way to change how long tasks retain context.

thsottiaux's announcement put the global reset at about 6 p.m. Pacific time and addressed it to all paid subscriptions. The follow-up confirmed completion.

Asked whether another reset was planned, thsottiaux gave no cadence.

The banked-reset documentation describes an account-held reset as a one-time item that expires and can be activated from the usage view. Separately, OpenAI’s Business release notes describe Premium seats as having 5x Standard capacity, no five-hour limit, and predictable weekly resets.

The evidence points to separate short-window and weekly meters, rather than a single token counter.

Reported measurements varied:

• A single 2.5-hour session consumed an entire limit for cto_junior, cto_junior's report said.
• cedric_chee recorded 300,000 input tokens, 50,000 output tokens, and 5.8 million cache-read tokens in roughly 30 minutes, or about 6 million tokens total, in cedric_chee's token accounting.
• The Developer Community report says its author saw a renewed five-hour allowance fall to 28% in about five minutes, reach zero about 30 seconds later, then exhaust a second reset after 3 minutes and 16 seconds.

OpenAI’s public guidance tied usage to reasoning effort, saying Astra on low outperforms the previous Sol model on high.

A separate reply put Astra low at about 50% cheaper per task than Sol high at roughly equivalent output, according to thsottiaux's comparison. The later 3-4x figure was specifically scoped to long-tail usage by Astra power users.

That scope is central to the disagreement. lateinteraction argued that tasks carrying more than 100,000 tokens of context could make Astra medium far more expensive than Sol xhigh, in lateinteraction's context reply.

Continuous agents make the subscription meter a concurrency problem. A rough estimate from nicdunz put one nonstop Astra Max thread at 3.7 days to drain a Pro 20x weekly limit.

• Max: 1 continuous thread
• XHigh: 1.1 threads for the same drain
• High: 1.13 threads
• Medium: 1.26 threads
• Lite: 1.75 threads

pvncher attributed Ultra’s aggressive draw to cache misses for every spawned subagent in pvncher's reply, a claim that would make parallel delegation more expensive than a single persistent thread.

Community posts point to an experimental context-management mode for overnight tasks. pvncher said it “makes all the difference” in pvncher's compaction report, then said it should be the default in pvncher's follow-up.

The post describes the mode as retaining notes and searching earlier history instead of repeatedly summarizing a conversation. It gave this configuration:

According to aibuilderclub_'s configuration post, the option is off by default, requires ChatGPT sign-in on Plus, Pro, or Pro Lite, and begins with a new task. Its claimed quota savings are user reports, rather than a guaranteed metering change.