This week, we launched our AI assistant powered by Bedrock agents using the Claude 4.5 model. On the first day, there was an outage in the EU region that AWS admitted to, and since then, we've encountered multiple spikes of ServiceUnavailableExceptions, even under low loads. We primarily use the EU models, although the global ones seem more stable but have higher latency. I'm curious about others' experiences with these popular models in Bedrock—are you running any production workloads on them? We're considering switching to the expensive provisioned throughput, but it appears those options aren't available for modern models, and the EU region seems to lag behind the US. How do you handle reliability with these setups?
4 Answers
I haven't found many situations where provisioned throughput is financially viable; it's just too pricey. We've been using the US region with cross-region inference on Sonnet 4, and it's been pretty stable lately—though it was hit or miss at first. If you have a Technical Account Manager, they might help you get in touch with the service team. It seems like there could be capacity issues in the EU, so you might want to think about falling back to US models even with the higher latency.
I'd suggest switching to Vertex for 4.5—it's much faster and more reliable from my experience.
If you’re feeling stuck with Bedrock, consider downgrading to Sonnet 3.7; it's not a huge difference. Also, Snowflake offers Sonnet options.
Never build your production agents to be fully reliant on Bedrock. It's useful, but always have a backup plan to switch to another provider or a self-hosted model. I currently have a failover setup that goes Bedrock -> direct Anthropic -> Ollama hosted model.
Related Questions
Neural Network Simulation Tool
xAI Grok Token Calculator
DeepSeek Token Calculator
Google Gemini Token Calculator
Meta LLaMA Token Calculator
OpenAI Token Calculator