OpenAI pauses new $200 Pro sign-ups
Existing subscribers keep access, but OpenAI gave no timetable for reopening the premium tier.
By Saeed Ezzati7 min read
The audio edition
Listen to this newsletter
0:003:35
Read transcript
OpenAI has paused new sign-ups and upgrades for its two-hundred-dollar-a-month ChatGPT Pro plan, after demand for GPT-6 Astra put pressure on its infrastructure. Existing customers keep access, but the company has given no timetable for reopening the tier. OpenAI describes this as targeted capacity control, not a withdrawal of Astra. Product leader Thibault Sottiaux said the two-hundred-dollar plan puts the greatest strain on the company’s systems, making it the narrowest point at which to limit incoming demand while preserving broader access. The one-hundred-dollar Pro plan, Go, Plus, and API access remain available. New customers cannot move up from Free, Go, Plus, or the one-hundred-dollar tier into the paused plan. There is also a sharp edge for current subscribers: anyone who cancels or lets the plan expire cannot buy it again until the pause is lifted. That makes the tradeoff unusually visible. OpenAI is giving up potential premium revenue to protect service quality for people already inside the most demanding tier. The company has not disclosed how many people were trying to join each day, or how long the constraint may last. So the operational signal is clear, while the scale is not. Astra demand is testing whether premium access can be expanded at the same pace as model capability. For founders and AI-heavy teams, the practical question is less about the headline price than about access stability: which users get continuity, which workloads remain available through the API, and how much capacity providers are willing to reserve before opening the gate again. That same tension between capability and control appears in Anthropic’s report on Claude. The company says it removed accounts linked to nine influence operations across six continents, involving governments, state-aligned media, private firms, political operators, and an exiled opposition movement. The reported activity included fake personas and news sites, forged documents, political content, and repeatable workflows using persistent files, source lists, evasion rules, and batch calls. One Russian state-aligned operation in the Central African Republic was assessed as Breakout Scale Category Four. Anthropic says it can sometimes detect planning before publication, but its direct view largely ends when material moves into wider media networks. That boundary matters: removing accounts can interrupt production, but not necessarily distribution. And in hiring, the control problem can begin before an AI system scores anyone. A University of Georgia-led study of hundreds of asynchronous job applicants found that people expecting AI review were more likely to exaggerate their qualifications. The tested AI rated embellished and honest candidates with equivalent qualifications about equally, while human reviewers generally favored authentic answers. Giving applicants specific information about the system’s criteria reduced exaggeration to levels comparable with expected human review. The result is limited to this study and tool, but it highlights a broader design issue: opacity can change the behavior being measured. A more constructive model of access comes from NASA and IBM, which released an open-source lunar AI model with code, datasets, and benchmarks. Trained on roughly two million image tiles from several missions, the NASA-IBM Lunar Foundation Model can be adapted for crater mapping, volcanic-feature detection, and estimates of where polar ice may be stable. NASA and IBM report especially strong results for ice-stability analysis, including a cited error reduction of up to twenty-two percent against a SwinV2-B baseline, though changing orbital lighting can make smaller craters harder to detect. Across these stories, the useful thing to watch is not just what models can produce. It is who controls access, what users adapt their behavior to, and whether the surrounding system can still see the result after it leaves the model.
