AgentCore gateway rate limits publish real numbers: 80,000 TPM for restricted models, 20 RPM per Basic user

AWS Machine Learning Blog·medium signalXBlueskyLinkedInCopy link

AWS shipped per-user, per-target rate limiting on the Bedrock AgentCore gateway on August 6, covering MCP targets, inference targets, and HTTP passthrough with three metrics — requests (RPS/RPM, all targets), tokens (TPM, inference only), and connections (CPS, all targets). Limits scope by JWT claims (`$.context.jwt.sub`, `.role`, `.azp`) or IAM identities (`$.context.iam.principal`, `.sourceIdentity`), and by `targetName`, `toolName`, or `qualifiedModelId`. The CLI examples give usable defaults: 100 RPM/50 CPS for Basic tiers, 300 RPM/150 CPS for Advanced, 20 RPM per individual, and token caps of 80,000/40,000/20,000 TPM by tier.