tomorrow 9:15am ET: the full dataset drops. 3 days of claude code traffic through a logging proxy - 6,673 requests, what actually served them, and the mechanism behind the silent reroute everyone argues about from theory instead of logs. it is more boring and more damning than the conspiracy.

i logged every model claude code actually served me. 7,128 requests in 3 days. 94.6% went through the fallback path. every direct attempt: rate-limited, zero exceptions. 60 seconds of the proxy catching it live:

1/ 94.08% of my agent traffic went through a fallback i picked myself. 6,673 requests over 3 days, the third one partial, every one logged by a proxy i run in front of the client. the log is more interesting than the pitch. thread:

claude code silently reroutes your requests to a different model. no banner, no warning. i run a 123-line proxy that logs every one and lets you pick the fallback. 7,128 requests in 3 days: 94.6% on the fallback path, every direct attempt rate-limited. receipts: