Tibo Sottiaux, head of OpenAI Codex, stated that the demand for Astra is "unprecedented," prompting the company to leverage all available resources to maintain supply. He noted that even during steeper growth, the current demand level has not been seen. He emphasized prioritizing the experience of existing users and mentioned that if demand continues to rise, there may be a temporary pause on new Pro subscriptions.

Astra refers to GPT-6 Astra, which opened to select institutions on September 3, and subsequently expanded to ChatGPT Plus, Pro, Business, Enterprise, as well as API, Azure, and Bedrock. Plus users primarily utilize Astra in Work and Codex, while the GPT-6 Pro in standard chat windows is mostly limited to Pro tier. The Pro tier, priced around $100, allows approximately 50 messages per week, while the $200 tier allows about 200 messages per week, showing a significant increase in usage compared to the previous Sol model.

Upon launch, Astra faced capacity bottlenecks. Some paying users have not been able to access the model, leading Tibo to announce that starting September 3, users would receive a reset of pre-stored usage for each day they were unable to access the model; Sam Altman described the launch as "messy." Subsequently, Astra was included in the regular quota, with claims that long-tail usage optimization could reduce deductions by three to four times. Codex's weekly active users had previously reached around 20 million. Users continue to report full capacity, rapid quota depletion, and that Astra consumes more quota than Sol for the same tasks.

The warning pertains to new Pro subscriptions, not the termination of existing ones. Tibo had previously promised to retain Codex on free and Plus tiers. Reports indicate that Astra's API pricing is about $10 per million inputs and about $50 per million outputs, higher than the replaced model. Training and inference are still tied to large-scale GPU clusters, with NVIDIA stating that the next batch of approximately 400,000 GPUs is on the way, but peak inference is determined by subscription influx.

Mechanically, this is a quota gate, not a rollback of model capabilities. The buyers are paying users seeking the strongest agents to write code and control computers; the seller is OpenAI, which is putting cutting-edge models into a fixed monthly fee. Pausing new Pro subscriptions effectively allocates incremental revenue to stabilize existing user experiences, avoiding a slowdown for all. The beneficiaries are existing Pro and enterprise seats; those left out are individual developers wanting to get on board during the launch week. The computing power has not been "used up"; it is simply allocated based on subscription priority.

Source: Public Information