Lower threshold: cheaper, more requests go to small models. Higher: safer, more go to Nova Pro.
Strands Decider
2B decision model on AgentCore Runtime (CPU)
Waiting for a request.
LLM classifier
Nova Micro, one-word answer
Waiting for a request.
LLM classifier
Claude Haiku 4.5, one-word answer
Waiting for a request.
Bedrock prompt router
Nova Prompt Router (Lite or Pro)
Waiting for a request.
Answer
The decider's chosen model answers here.
0
your requests
$0
your cost, routing included
-
than always Nova Pro
0
answers served by the playground
Loading benchmark results.
Quality against cost
Added latency of the router (median)
Every strategy on the held-out test set
Quality by task family (test set, original prompts)
What happens when you press the button
- Three routers look at your request at the same time. The Strands Decider answers one typed question, which is the cheapest model that will answer this correctly?, with a probability for Nova Micro, Nova Lite and Nova Pro. The page routes to the cheapest model whose cumulative probability of being enough reaches the threshold.
- Nova Micro and Claude Haiku 4.5 get the same question and the same model descriptions and answer with one word.
- Bedrock Intelligent Prompt Routing (the default Nova router) picks Nova Lite or Nova Pro and answers in the same call.
- The decider's chosen model answers your request. The savings figure compares its cost with Nova Pro answering the same tokens.
Limits
No sign-in and no key. Each visitor gets a fixed number of calls per hour, the whole playground has a daily cap, prompts are limited in length, and answers in output tokens.
The decider runs on 2 vCPUs in a microVM, so it takes several seconds per decision; when it is busy with another visitor the card says so. Source, benchmark and analysis: github.com/Vivek0712/strands-decider-agentcore.