When I create a custom Opus 3 assistant and enable internet access, the response outputs a <thinking> block, which appears to execute prior to my custom preamble. For example, here is the output using the custom assistant with the custom preamble "don't use vowels":
This is a recent change, because I've had a custom assistant using Opus 3 and internet access for several months, and it never output the <thinking> block. This breaks immersion and also leads to much slower generation of the actual answer. This also doesn't seem to happen with other models, like Sonnet 3.5 or GPT 4o, nor does it happen when I disable internet access for Opus 3.
