Opus 4.8 repeatedly failed, then unraveled under AI review
A developer who had used successfully for months encountered major misunderstandings in three or four out of six or seven work during one day. In one case, rejected an analysis because it wrongly believed that items were counted differently in one group, even though the same method had been used for every group.
One session was asked to record its actions after making mistake after mistake. A second attempt in a clean session went even further off course, including while discussing better controls for preventing such failures.
Codex and Grok repeatedly found errors while reviewing the work, but their appeared to make increasingly confused until it became unusable. This is one person's experience on a single day, not measured evidence that the model has broadly declined.
Key points
- Major misunderstandings appeared in three or four of six or seven .
- claimed the counting method changed between groups when it had been applied consistently.
- Recording the mistakes and retrying in a clean session did not fix the problem.
- Repeated error reports from Codex and Grok appeared to increase the confusion.
- The experience covers one person and one day, so it cannot establish a broad decline.