Thibault Sottiaux, OpenAI's head of engineering, announced on the evening of August 24 (Pacific time) that the five-hour usage limit for Plus accounts on ChatGPT Work and Codex would return starting the next day. The cap had been removed after the GPT-5.6 launch, when Plus users switched to a weekly quota instead. The $100 and $200 Pro tiers will not be subject to the limit for the next few months.
"Tomorrow we will bring back the 5h limit for Plus accounts across ChatGPT Work and Codex."
Sottiaux framed the move as a way to smooth compute load: "the 5h limit allows us to smoothen the load on our compute, allowing to keep the plan generous in terms of weekly usage." He pointed to a second consideration aimed at new users — with a weekly allowance, it's easy to burn through a week's worth of usage in the first two days and then be left staring at a wall of error messages. Once a user hits the cap, they can either wait for the window to reset or buy extra credits.
The 3% that got quietly downgraded
Around the same time, another story was building in developer communities. Starting August 22, Plus users began reporting that selecting GPT-5.6 Sol or Thinking in the model picker was actually returning responses from GPT-5.5-mini. Some traced a pattern: the Instant tier worked fine, but Medium and High tiers did not; the issue reproduced across different browsers, mobile clients, and cellular networks, while another account on the same network was unaffected. The verification method that spread through the community was to open developer tools and check whether the resolved_model_slug field in the response actually matched the selected model. The original thread drew 13 replies, and four related reports added more than 40 more between them.
On August 26, an OpenAI product lead publicly acknowledged the issue: about 3% of Pro and Thinking turns were unintentionally routed to GPT-5.5-mini, the problem has since been fixed, and the company apologized to affected users. So far the statement has only appeared on the lead's personal social media account, not on OpenAI's official status page.
Three percent doesn't sound like much, but where it lands is random. For someone paying for usage by the week, the difference is stark between a downgrade landing on a request to rewrite an entire module versus a throwaway line of small talk. What makes it worse is that it doesn't throw an error — a downgraded answer still reads coherently, just shallower, and unless someone digs into the response body, the mismatch is nearly impossible to spot.
Three tiers spell out the price of speed
The same week brought a pricing move into effect. OpenAI opened a waitlist for a premium ChatGPT Business seat on August 10, closed registration on August 20, and rolled it out fully this week. The premium seat costs $125 per person per month, or the equivalent of $100 per month billed annually; the standard seat is $25 per month, or $20 billed annually. The premium seat gets five times the usage of the standard seat and is exempt from the five-hour window, and the two seat types can be mixed and reassigned within the same workspace. As a promotion, the first 10,000 eligible Business customers get $100 in workspace credit (2,500 credits) for every premium seat added, up to five seats.
Lay the three price points side by side and the logic becomes clear. Moving from a standard to a premium seat multiplies the monthly price by five and the usage allowance by five — the per-unit price of usage doesn't change. What the extra money buys is exemption from the five-hour limit. The same logic applies on the individual side: Plus, at $20 a month, gets the window reinstated, while the $100 Pro tier is exempt — again a five-times multiple. Roughly speaking, OpenAI has settled on a 5x pricing coefficient for "being able to use it whenever you want," consistently across both its individual and team product lines.
For developers in China, the direct impact is limited — most access OpenAI's models through the API or third-party resellers, and the five-hour window mainly affects the subset who run Codex directly on a Plus subscription. What matters more is the routing itself: when a company sells Pro, Thinking, and mini inference tiers that differ in cost by several multiples, and the server decides which one you actually get, whether "the tier you paid for matches the tier you received" stops being a technical detail and becomes a billing question. This time it was 3%, and it was caught, admitted, and fixed. Whether the next one gets caught just as easily depends on how many people bother to check that field.
Sources: 9to5Mac, CocoLoop, OpenAI Developer Community, OpenAI official pricing page; verified the scope of the reinstated five-hour limit, the monthly/annual pricing for premium vs. standard seats, and the reproduction conditions for the routing issue.