Bedrock Quotas

For raw quota documentation, use the AWS MCP Server’s aws___search_documentation / aws___read_documentation tools or see: Quotas for Amazon Bedrock (opens in a new tab) | How tokens are counted (opens in a new tab)

Why This Matters for Migrations

OpenAI and Gemini APIs have high default rate limits for paying customers (OpenAI Tier 4/5: 2M+ TPM). Bedrock default quotas start lower and scale with account history. The quota risk assessment below works without an AWS account — it uses the user’s current token volume and selected model to flag whether a quota increase will be needed, so teams can plan the request into their migration timeline.

The 5× Burndown Trap

Claude models (3.7+) on Bedrock consume quota at 5× the rate for output tokens. This is the single most common surprise for teams migrating from OpenAI (which has no output multiplier).

Effective TPM by workload type (at 200K TPM quota on Claude):

Formula: Effective TPM = TPM_quota / (input_ratio + output_ratio × 5)

Nova, Llama, DeepSeek, and Mistral models have 1× burndown (no multiplier). Output-heavy workloads that hit Claude quota limits may benefit from routing to these models for applicable tasks.

Quota Risk Assessment

Apply during Design phase after model selection. Uses ai_token_volume from preferences.json and the selected model family.

Include quota_risk in aws-design-ai.json → ai_architecture.

max_tokens Reservation Trap

Bedrock deducts max_tokens from TPM quota at request start, before any tokens are generated. Unused quota is returned after response completes.

If code sets max_tokens: 4096 (common OpenAI default) but typical responses are 200 tokens, each request reserves 20× more quota than needed — dramatically reducing concurrency.

Migration action: Set max_tokens to ~1.5× expected output length. This is a code change that should be flagged in the migration guide when quota_risk is "medium" or "high".

Cross-Region Inference as Quota Mitigation

Cross-region inference (CRIS) profiles have separate, higher quotas because traffic distributes across multiple regions. For production migrations with quota_risk = "high":

• Recommend CRIS as the default deployment mode
• No code changes required (use the CRIS inference profile ID instead of the regional model ID)
• Trade-off: slightly higher latency variance, ~10% price premium in some regions

Pre-Migration Quota Checklist

Surface these items in the production readiness checklist (generate phase) when quota_risk ≥ "medium":

• Not done: Quota risk assessed (this can be done before having an AWS account — it’s based on token volume and model selection)
• Not done: AWS account created and Bedrock model access enabled in target region
• Not done: Current TPM quota checked via Service Quotas console for target model + region
• Not done: Quota increase requested if peak TPM exceeds default (allow 1–5 business days)
• Not done: max_tokens set to ~1.5× expected output (not model maximum)
• Not done: Cross-region inference profile evaluated for high-volume workloads
• Not done: Retry configuration uses adaptive mode (botocore.config.Config(retries={"mode": "adaptive"}))
• Not done: CloudWatch alarm configured on ThrottlingException metric > 0

How to Request a Quota Increase

Include this guidance in migration artifacts when quota_risk ≥ "medium". Users may not have an AWS account yet when the plugin assesses risk — these instructions tell them what to do once they do.

Where: AWS Console → Service Quotas → Amazon Bedrock → search for the model name (e.g., “Claude Sonnet 4” or “tokens per minute”)

What to request:

• Find the quota named like: Tokens per minute for [Model Name] or Cross-region model inference tokens per minute for [Model Name]
• Click “Request increase at account level”
• Enter desired value based on the effective TPM calculation above

How to calculate the target value:

Where burndown_multiplier = 5 for Claude models (because output tokens consume 5× quota), 1 for all others. If the user doesn’t know their peak TPM, estimate from daily volume:

Timeline: Allow 1–5 business days for approval. Request increases during the setup/provisioning stage of migration, not during cutover.

Alternative — cross-region inference: If the increase isn’t approved in time or the requested value is very high, cross-region inference profiles provide higher aggregate capacity without a quota increase request (traffic distributes across regions automatically).

Integration Points