As of September 7, 2026. Short answer: Astra can draw subscription usage faster than Sol depending on the task, reasoning level and Fast mode, and Astra Fast is billed at 2.5× Standard. On September 7, OpenAI cut long-tail usage by up to 3–4×, but it was not announced as a refund of quota already spent. A full banked reset, by contrast, refreshes both the five-hour and weekly Codex windows. This is not a story about Astra being weak. It is almost the opposite. Astra is strong and fast, so people use it. Then the usage meter falls through the floor. After that, the operator announces a major efficiency fix. The natural response is: great—so what about the fuel we already burned?
1. What happened: strong → fast → quota gone → efficiency fixed → quota already gone
On September 7, OpenAI's Tibo offered a blunt calibration for reasoning effort: GPT-6 Astra on Low performs better than GPT-5.6 Sol on High, and users who were satisfied with Sol High should try Astra Low or Medium.
Low, Medium, and High are not absolute intelligence grades shared across models. Astra Low does not mean “dumb Astra.” It means the newer engine can be run at a lighter reasoning setting and still clear a bar that required Sol to run at High.
Naturally, users hear this and think: then Astra Low should be the default—stronger, faster, done. That logic is mostly sound. The surprise was the fuel gauge.
2. “Capacity” and “subscription usage” are two different things
If capacity means context size, Astra is not simply larger than Sol. OpenAI lists 1.05 million tokens of context and 128,000 maximum output tokens for both Astra and Sol. The container is effectively the same size.
Billing treatment is not identical, however. OpenAI's Enterprise token-based rate card normally applies additional multipliers beyond 272K input tokens for long context, but it explicitly gives GPT-6 Astra in Codex an exception from those extra long-context multipliers above 272K. That is not extra physical context capacity; it is different metering for using a very large context.
The five-hour and weekly subscription allowance is another meter. OpenAI explicitly says Astra can consume allowance faster than Sol depending on the task, input/output size, reasoning level, and Fast mode.
3. Why Astra uses so much: per-token price and Fast mode
Public API prices show the compute gap: Astra is $10 per million input tokens and $50 per million output tokens; Sol is $4 and $20. That is 2.5× per token. But API prices are not a direct formula for the percentage shown on every ChatGPT subscription meter.
One reply to Tibo reported that a job normally costing about 10% of the weekly allowance on Sol Medium took about 50% on Astra Low. That is one user's anecdote, not a universal benchmark. Context length, tools, caching, reasoning and execution mode can all change the result.
Another user said efficiency became much better after disabling 2x speed. There is a real mechanism behind that observation. OpenAI's current Work/Codex rate card charges GPT-6 Astra Fast at 2.5× the Standard rate. Speed is not free. If you floor the accelerator and then stare suspiciously at the fuel gauge, the rate card is quietly staring back.
4. “Fast” matters for more than vibes—Tibo says Astra moved internal plans six months forward
If Astra were merely a shinier model with a worse fuel bill, people could simply return to Sol. But on September 5, Tibo said Astra had been one of OpenAI's biggest competitive advantages before general availability and that the productivity jump was large enough to move some plans from the middle of next year to DevDay, roughly six months earlier.
That is an internal OpenAI claim, not an independently audited productivity study. Still, it explains why wall-clock speed matters. Agentic work is not just one-answer benchmark accuracy. It is also how quickly long jobs finish, how often they retry, how well they retain context, and how many complete tasks fit into a working day.
Astra's appeal can therefore be larger than “one more benchmark point.” It may change how many finished jobs fit into the same day. That makes users want to run it. And running it is exactly how the quota disappears. A beautifully engineered trap.
5. The September 7 fix: up to 3–4× less usage in the long tail
A few hours after the effort calibration, Tibo announced another change aimed at power users of Astra who are logged in with their ChatGPT account, specifically the long tail where heavy or long-running work can draw unusually large usage.
The promise was unusually clean:
- no quality change;
- better usage efficiency in the long tail;
- in affected cases, up to 3–4× less subscription usage drawn.
This does not mean every task instantly costs one quarter as much. It is an up-to figure for long-tail cases. But for long sessions it is a substantial change.
For illustration only: if the anecdotal 50% task were fully covered by a 3–4× reduction, the arithmetic would land around 12.5–16.7%. That is not a prediction or guarantee. It simply shows how large the announced improvement is relative to the complaint.
6. Better future mileage does not refill yesterday's tank
Put the timeline in order:
- Astra arrives.
- Users are told Low already beats Sol High.
- It is fast, so they run it hard.
- The weekly meter evaporates.
- OpenAI says long-tail usage can now be 3–4× more efficient.
- Excellent. The tank is empty.
The September 7 change improves future draw. It was not announced as an automatic refund of subscription usage already consumed before the fix.
So a technically excellent fix produces a psychologically obvious response: please give back the fuel the old meter already drank.
7. How banked Codex resets work, and why users keep asking for one
There is recent precedent. On September 5, Tibo announced a full banked reset for Plus, Pro and Business users as part of the Astra rollout celebration. OpenAI's help documentation says redeeming a full banked reset refreshes both the five-hour and weekly Codex windows and moves the weekly reset date.
A banked reset is not a permanent allowance increase. It is a one-time promotional reset, and OpenAI explicitly says future resets are not guaranteed. In the latest Tibo feed checked late morning Japan time on September 7, the new posts covered the Low/Medium guidance and the 3–4× long-tail improvement, but no additional reset after the September 5 announcement was visible.
So the scene writes itself:
Tibo: We fixed efficiency.
Users: Good.
Tibo: No quality loss.
Users: Excellent.
Tibo: Up to 3–4× less usage in the long tail.
Users: Fantastic. And yesterday's weekly quota?
Tibo: …
Users: Tibo. Reset. Please.
8. Which setting to use: Astra Low + Standard now looks like the best default
A sensible operating policy, based on the current information, is straightforward:
- Heavy everyday work: Astra Low + Standard. Aim for Sol-High-class capability without unnecessary reasoning and Fast overhead.
- Hard jobs that actually fail on Low: Astra Medium. Escalate when needed instead of pinning everything to High.
- Latency matters more than quota: Fast. Astra Fast is billed at 2.5× Standard in the current Work/Codex rate card.
- Economics and allowance longevity matter most: Sol still matters. Astra being smarter does not erase Sol's lower unit cost.
The important lesson is not to compress capability, speed, context size and quota efficiency into one word called “better.” Astra can be stronger and faster while having the same context capacity as Sol and a heavier usage profile. Codex gives Astra a special long-context metering exception, and the September 7 patch appears designed to cut the worst long-session cases substantially.
Which leaves one final, deeply human request:
We understand the mileage is better now. Please let us test it with a full tank.
References (10)
- OpenAI Models: https://developers.openai.com/api/docs/models
- GPT-6 Astra model card: https://developers.openai.com/api/docs/models/gpt-6-astra
- ChatGPT Work and Codex: https://help.openai.com/en/articles/20001275
- ChatGPT rate card: https://help.openai.com/en/articles/11481834
- Enterprise token-based rate card: https://help.openai.com/en/articles/20001415
- How banked Codex resets work: https://help.openai.com/en/articles/20001498
- Tibo — Astra Low vs Sol High: https://x.com/thsottiaux/status/2096688770523467947
- Tibo — long-tail usage improvement: https://x.com/thsottiaux/status/2096717905614524491
- Tibo — Sep. 5 full banked reset: https://x.com/thsottiaux/status/2096035437299237298
- Tibo — Astra and six-month productivity shift: https://x.com/thsottiaux/status/2096101429832552872
