Key Info

Requesty's measurements show claude-opus-5-5 p50 latency varies sharply by endpoint: vertex@us at 1.32s versus bedrock at 17.9s. The fastest endpoint changed 13 times in a single 24-hour window.

Highlights

  • claude-opus-5-5 p50 latency: vertex@us 1.32s, bedrock 17.9s
  • Over 13x latency difference between endpoints serving the same model
  • The fastest endpoint rotated 13 times in 24 hours
  • Pinning a single endpoint means inheriting its worst-performing hours