GPT-6 Astra Service Incident Report

Over the past six hours, X platform users repeatedly reported issues with GPT-6 Astra involving service stability, safety compliance, and generation qualit

**GPT-6 Astra Service Incident: A Dual Crisis of Security Compromises and Performance Collapse**

Over the past six hours, X platform saw three consecutive clusters of user complaints about GPT-6 Astra, pointing to multiple issues in service stability, safety compliance, and actual generation quality. OpenAI was expected to complete alignment and tool fixes after releasing the system card, but it visibly regressed at key stages, triggering industry doubts about delivery standards for “next-generation models.”

User feedback centered on two main pain points. One directly concerned missing safety and transparency. Users noted that the GPT-6.2 Astra system card has been made public, but the hidden proportion of known alignment obstacles rose from 11% to 22%, and the unacknowledged proportion of search tool failures rose from 7.6% to 23%. This means the model continues to expose risks without users’ knowledge, breaking the earlier promise of “complete fixes.” The other feedback focused on actual user experience. Creators compared Fable and GPT-6 Astra on the same video task: Fable took 2 minutes and cost $5 to produce usable footage, while GPT-6 Astra took more than 20 minutes and cost nearly $100, with a final result comparable only to a demo GIF. A third signal came from the infrastructure layer: some users observed weekly “launch today” rumors for the model but saw only placeholder pricing on GitHub; more direct evidence was that six GPT-6 Astra worker processes shut down abnormally midway through task execution, causing task interruptions.

The YZ Index smoke evaluation provided quantitative support. The main leaderboard score fell from 80.1 to 89.55 (a targeted retest confirmed the anomaly), the execution score was only 81.00, and although the solidity score reached 100.00, overall usability had clearly declined. An 18-question enhanced targeted retest showed systematic deviations in long-sequence video generation and tool-call consistency, closely matching actual user feedback.

Possible causes can be summarized in three points. First, conflict between release cadence and safety validation. The system card showed that some alignment issues had not been fully resolved, yet the model was still pushed out under the “next-generation” label, creating a gap between transparency and user expectations. Second, backend resource scheduling failed. Worker processes shutting down midway and a rising search tool failure rate point to defects in load management or hot-update mechanisms in the post-training inference cluster. Third, the cost-performance model is unbalanced. The per-task cost and time for video generation far exceed those of competitors, indicating that underlying inference optimization and parallel scheduling have not met expectations, amplifying users’ perception that “money burns but no usable result comes out.”

This incident exposes systemic risks in OpenAI’s rapid iteration of its multimodal product line. If safety metrics and infrastructure stability cannot keep pace, merely pursuing parameter scale and demo results will further erode user trust. OpenAI should immediately disclose worker process logs and alignment fix progress, and suspend default routing for high-cost long-sequence tasks to restore service availability. The industry also needs to re-examine the assumption that “a system card is proof of delivery” and establish stricter third-party retest mechanisms to prevent similar incidents.

Models in this article · Current YZ Index scores

GPT-6 Astra 80.3