Amazon Bedrock On-Demand vs Provisioned Throughput: Unit Economics Breakdown and Model Decision Tree

Amazon Bedrock Model Menu Overview (Nova, Anthropic, Meta, and More) Amazon Bedrock provides access to over 100 foundation models (FMs) from leading providers, all via a unified API. This serverless platform supports enterprise AI stacks like LUMOS, enabling Retrieval Augmented Generation (RAG), agents, and custom fine-tuning without managing infrastructure. Key providers and flagship models (exact model ids from AWS docs as of May 15, 2026 UTC): - Amazon Nova family : - (high-capability reasoning and multimodal) - (balanced performance) - (cost-efficient for lighter tasks) - (ultra-low latency) - Specialized: (creative), (video), (speech) - Anthropic Claude : - (frontier reasoning) - (versatile) - (speed-focused) - Meta Llama : - (open weights powerhouse) - (efficient scaling) - Others: Models from Mistral, Stability AI, and the Bedrock Marketplace for niche deployments. Bedrock's model

menu evolves rapidly, with features like Guardrails, latency-optimized inference, and cross-region options. Always verify latest availability at . On-Demand vs Provisioned Throughput: Key Differences Bedrock offers two primary inference modes: On-Demand and Provisioned Throughput . - On-Demand : Pay-per-use based on input/output tokens processed. Ideal for variable, bursty workloads. No commitments—scale instantly with SLAs for latency and availability. Supports all models dynamically. - Provisioned Throughput : Reserve dedicated model units (MUs) for 1, 6, or 12 months. Each MU guarantees throughput (e.g., tokens per minute) at lower per-token rates. Suited for predictable, high-volume enterprise apps like RAG pipelines or agents. Minimum commitment: 1 MU per model. Key diffs: - Cost structure : On-demand is flexible but higher unit economics; provisioned offers 30-70% savings on token

s for committed volume (per AWS pricing methodology). - Scalability : On-demand auto-scales; provisioned provides reserved capacity. - Latency : Provisioned often lower/consistent for steady loads. - Customization : Provisioned supports custom models; on-demand is foundation models only. See for details. Unit Economics Breakdown: Pricing Impacts by Model and Mode Unit economics hinge on tokens processed, mode, region, and volume tiers. Calculate as: Total Cost = (Input Tokens × Input Rate) + (Output Tokens × Output Rate) + Fixed Fees (provisioned only) Factors: - Token multipliers : Images/videos count as extra tokens (e.g., 1 image 1K tokens for Nova multimodal). - Batch discounts : Up to 50% off for non-real-time jobs. - Tiers : On-demand has volume discounts; provisioned pricing per MU/hour + per-token overage. Provisioned shifts economics: Pay upfront for capacity, then discounted to

kens. For a RAG app with 10M daily tokens, provisioned can cut costs 50%+ vs on-demand (methodology from AWS calculator). Use AWS Pricing Calculator for simulations: . Official Pricing for Top Models (As of May 2026) Pricing per AWS Bedrock docs at (US East 1, as of May 15, 2026 UTC). Always confirm current rates—SKUs update frequently. On-Demand Examples (per 1M tokens): - : Input $15, Output $75 (high-end reasoning). - : Input $3, Output $10 (optimized multimodal). - : Input $5, Output $20. Provisioned Throughput : Hourly MU rate (e.g., $20-100/MU depending on model) + reduced tokens (50-90% off on-demand). For Claude Opus: 1 MU ( 500 tokens/min output) at $85/hour commitment, tokens at $3 input/$15 output. Sample Monthly Estimate (enterprise RAG: 1B input/200M output tokens): - On-Demand: $20K (Nova Pro). - Provisioned (6-mo, 10 MUs): $8K + commitment ( $50K total, but 60% unit saving

s). Cross-region inference adds 20% premium. Batch API: 25-50% off. Decision Tree: Picking the Right Model Profile for Your Workload Use this markdown decision tree for Bedrock selection in LUMOS-like stacks: Steps: 1) Profile workload (tokens, latency SLA). 2) AWS Cost Explorer sim. 3) Test via Bedrock console. Scalability Tradeoffs: Latency, Throughput, and Cost Optimization - Latency : On-demand p95 <1s (Haiku); provisioned <500ms guaranteed. - Throughput : Provisioned scales to 10K+ tokens/sec per MU cluster. - Optimization : Use quantization, prompt caching (up to 90% cache hit savings), In-Region inference (-20% latency). Hybrid: On-demand dev/test, provisioned prod. Monitor via CloudWatch. Enterprise Use Cases: RAG and Agents on Bedrock RAG : Nova Lite provisioned for doc search (low cost, 128K context). Monthly: 500M tokens → $2-5K provisioned vs $10K on-demand. Agents : Claude O

pus for tool-calling chains. Provisioned for 24/7 ops, saving 40% on multi-turn. Integrate with LUMOS: Bedrock as inference layer, Knowledge Bases for RAG. Migration Tips and Best Practices - Start with on-demand PoC, migrate to provisioned post-benchmark. - Tag resources for cost allocation. - Enab