I’ve noticed the same, especially with long-context window retrieval. It seems like the "needle in a haystack" performance has degraded slightly in 4.6 compared to the early 4.5 builds.

I'm getting more frequent "hallucinated constraints" where it refuses to follow specific formatting instructions that it used to handle perfectly. Are you seeing this more in the API or the web console? I’m wondering if it’s a latent quantization issue or just heavy traffic throttling.