Benchmark notes: Opus5 & Opus5.5 on us-east-1

I’ve been running load tests on opus5 and opus5.5 deployments in us-east-1 and want to share my observations.

Test setup

• Region: us-east-1
• Model: opus5, opus5.5
• Test type: Concurrent chat completions, mixed short and medium-length prompts
• Auth: Standard AWS Bedrock credential flow

Observations

1. For short prompts: single-request latency is roughly 3.0s ~ 4.5s. Longer prompts with large output targets can extend to 6s+.
2. Throttling (429 responses) triggers quickly once concurrency increases. Even with higher baseline TPM/RPM quotas, the opus family is far more sensitive to load compared to older models.
3. I’ve encountered a separate static 401 issue on some deployments: credentials pass initial auth checks, but opus5/5.5 model access fails, while other standard Bedrock models on the same subscription work normally. This appears unrelated to concurrency pressure.
4. The stable maximum throughput varies significantly between different provisioned-throughput deployments for the same model version.

Questions
I’m curious if other folks have run similar benchmarks for these models.
What steady throughput and latency are you getting on your working opus5 / opus5.5 deployments?
If you have a stable production deployment for these models, I’d be glad to exchange benchmark data and configuration details via DM.