Live log data from BasicallyMythos, a 123-line local proxy that observes every Claude Code model reroute. Updated continuously.
When Claude Code's safety layer flags a request, the request does not fail. It is re-issued to a different model, with no banner and no warning. This page documents what that looks like in real traffic.
Across all three days, every direct POST to the frontier endpoint returned 429. The fallback path (an open-weights model, Kimi K3) answered 98.2% of the 6,744 rerouted requests. The tool did not degrade gracefully; it failed over, silently, thousands of times.
[2026-07-21T09:01:57Z] anthropic model=claude-fable-5 POST /v1/messages -> 429 [2026-07-21T09:02:10Z] KIMI model=kimi-k3 POST /v1/messages -> 200
| Date (UTC) | Requests | Rerouted |
|---|---|---|
| 2026-07-19 | 677 | 574 |
| 2026-07-20 | 1,630 | 1,571 |
| 2026-07-21 | 5,223 | 4,973 |
| 2026-07-22 | 1,516 | 1,347 |
| 2026-07-23 | 2,069 | 1,959 |
| 2026-07-24 | 111 | 4 |
| 2026-07-25 | 219 | 3 |
| 2026-07-26 | 127 | 2 |
| 2026-07-27 | 27 | 0 |
| 2026-07-28 | 363 | 354 |
| 2026-07-29 | 1,180 | 1,180 |
| 2026-07-30 | 1,617 | 1,614 |
| 2026-07-31 | 2,631 | 2,631 |
| 2026-08-01 | 2,414 | 2,413 |
| 2026-08-02 | 410 | 378 |
| 2026-08-03 | 25 | 6 |
The reroute is client-side. When a safeguard fires, the CLI re-issues the request using the model id in ANTHROPIC_DEFAULT_OPUS_MODEL, to the same ANTHROPIC_BASE_URL as everything else. The downgrade is not a server decision you can appeal. It is a config value on your own machine, and a config value on your own machine is something you can see and choose.
BasicallyMythos is MIT-licensed, zero dependencies, loopback-only. Setup is about ten minutes. Run node reroute-stats.cjs for your own report. If your intercept rate looks like this, the author wants to hear about it: @TEA_Resistance.
Related coverage: GitHub issue anthropics/claude-code#19468 (systematic model degradation reports). This dataset is independent user-side observation, not affiliated with Anthropic.