Amazon Bedrock Cost Optimization: Routing Tasks Across Nova Micro, Lite, and Pro

What You'll Learn

• How to build a routing framework for Amazon Nova's three model tiers
• The classification prompt that assigns each task to the right tier automatically
• How to audit your existing Bedrock workloads in one hour
• The monthly cost savings calculation for a typical enterprise deployment
• Common routing mistakes and how to avoid them

Most organizations using Amazon Bedrock route all tasks to a single model. They picked a model that produces acceptable outputs, set it as the default, and moved on. This works. It also means they are paying Nova Pro prices for tasks that Nova Micro would handle just as well at roughly 23 times lower cost per token.

The Amazon Nova model family arrives with pricing that makes task routing worth building: Nova Micro at $0.035 per million input tokens, Nova Lite at $0.06, Nova Pro at $0.80. The right routing framework assigns each task to the appropriate tier based on complexity. Lower costs on standard tasks, full quality preserved on complex ones.

This Productivity Gem builds that routing framework from scratch, including the classification prompt that automates tier assignment, the audit process for existing workloads, and the savings calculation that justifies the implementation work.

SUBSCRIBER BREAK -- Premium Content Below

Tier Assignment Logic

Nova Micro (default for structured text tasks): Input is structured text; output is a structured response (classification, extraction, yes/no, named fields); no creative judgment required; output under 500 words. Nova Lite (default for multimodal and summarization): Input includes images, video, or mixed-format documents; task requires synthesis across a moderately long document; output is a summary or multi-point analysis. Nova Pro (default for complex reasoning and high-stakes outputs): Multi-step reasoning across ambiguous information; input spans multiple documents; output used to make a significant business decision; code generation or debugging.

The Classification Prompt

Use this as a preprocessing step to assign incoming tasks to the correct tier. Run it using Nova Micro itself: at $0.035/million tokens, the classification cost is negligible.

Classify the following task into the correct Amazon Nova processing tier. Tier definitions: - MICRO: Simple extraction, classification, or yes/no decision from clearly structured input. No judgment required. Short output. - LITE: Summarization, image or document analysis, moderate synthesis. Structured input, multi-point output. No complex reasoning. - PRO: Complex reasoning, multi-document synthesis, code generation, high-stakes output, or tasks where error cost is high. Task to classify: [DESCRIBE THE TASK IN 2-3 SENTENCES] Return only one word: MICRO, LITE, or PRO.

Auditing Your Existing Bedrock Workloads

• Export Bedrock usage data. In AWS Cost Explorer, filter by service = "Amazon Bedrock" and group by "API Operation" and "Model ID." Export the last 3 months.
• Identify your top 5 workloads by token consumption. Sort by total tokens consumed. The top 5 account for 80 to 90% of cost in most organizations.
• Classify each workload by tier. For each of the top 5, determine: what tier does this task belong to based on the criteria above? What tier is it currently running on?
• Calculate the savings opportunity. For each misrouted workload: monthly cost at current model vs. cost at the appropriate Nova tier.
• Prioritize by savings opportunity. Start with the highest-savings workload where tier reassignment carries the lowest quality risk.

Monthly Savings Calculation

Sample: 50 million input tokens per month, currently all routed to a mid-tier model at $0.25/million = $12,500/month. After routing: 40M tokens to Nova Micro ($0.035) = $1,400; 7M tokens to Nova Lite ($0.06) = $420; 3M tokens to Nova Pro ($0.80) = $2,400. Total: $4,220/month vs. $12,500/month. Monthly savings: $8,280. Annual savings: $99,360. The ratio shifts with usage volume, but most Bedrock workloads run on a higher tier than necessary.

Common Routing Mistakes

Underclassifying summarization: Long-document summarization needs Lite or Pro, not Micro. Routing to Micro produces truncated summaries. Overclassifying extraction: Invoice extraction, entity recognition, and field-level extraction from structured documents are reliably handled by Micro. Routing these to Pro is the most common source of unnecessary cost. Skipping fallback logic: Without fallback, a task that routes to Micro and produces a low-quality output produces a silent failure. Always implement fallback and log events. Not re-evaluating quarterly: As Nova models improve, tasks that required Pro may become reliable on Lite. Review tier assignments quarterly and test downward routing.