Source: Amazon Bedrock — How Tokens Are Counted Against Quotas

How Bedrock reserves `max_tokens` up front and applies output-token burndown multipliers to TPM/TPD quotas.

Contents

Summary

Explains the throttling arithmetic behind Bedrock’s tokens-per-minute (TPM) and tokens-per-day (TPD) quotas, with worked examples.

Key claims

  • At request start Bedrock deducts input tokens + max_tokens from TPM/TPD; throttling happens if that exceeds quota.
  • At the end it settles to input + cache-write + output × burndown rate and refunds the rest; cache reads do not count.
  • Burndown rates (late 2026): 15× for Claude 4.8; 10× for Claude Opus 5/5.5, Sonnet 5, Fable 5.1/5.2 and OpenAI GPT-5.6 models; 2.5× for Claude Mythos 5.x; 5× for Claude 4.7 and older; 1× otherwise unless the model card says so.
  • You are billed for actual tokens, not burndown-weighted ones (example: 1,000 in + 100 out at 5× uses 1,500 quota but bills 1,100).
  • An oversized max_tokens (32K for a 1K reply) cuts concurrency because the up-front reservation is large; size it near real output lengths using CloudWatch.
  • Models served only on the bedrock-mantle endpoint have separate input/output quotas and no burndown.

Feeds

Amazon Bedrock

Suggest changes →