After Astra's stealth nerf last night, we really need benchmarks to do a re-bench 1 week after any model release. This is ridiculous.

A user on reddit.com expressed frustration over a "stealth nerf" to the Astra model, claiming its capabilities significantly diminished overnight. They noted that Astra, which previously excelled at "1 shotting AAA game level models," now produces "completely terrible" results even after "reprompting on xhigh effort several times." The user criticized "frontier labs" for quietly reducing model capabilities to save compute, especially after many users purchased "Pro subscriptions" based on initial performance, suggesting the "FTC should get involved."

This report highlights a user's direct experience of a model's performance degradation, unlike many discussions that focus on theoretical capabilities or benchmark scores.