Optimizing Bedrock Costs

Optimizing Bedrock Costs

Generate prioritized, actionable Bedrock cost optimization recommendations backed by real Cost Explorer data.

Quick Start

Run the full optimization analysis:

bash ./_skills/custom/optimizing-bedrock-costs/scripts/optimize.sh [--days N]

Default: 90 days. Outputs TSV sections for each optimization area.

Optimization Areas

1. Model Tiering (Opus to Sonnet/Haiku Downgrade)

Opus models cost 5x more than Sonnet per token. Many workloads can run on Sonnet with equivalent quality.

Detection: Compare Opus vs Sonnet spend per region. If Opus > 30% of total model spend, flag for review.

Savings estimate:

Potential savings = Opus_cost * 0.80  (assuming 80% of Opus tasks can move to Sonnet)
Conservative savings = Opus_cost * 0.50

Pricing reference (USD per 1M tokens):

2. Cache Efficiency

Target: >60% hit ratio, >4x reuse ratio.

Detection: Pull token-level usage types, compute hit ratio and reuse ratio per month.

Recommendations by finding:

Savings estimate:

If hit_ratio improves from X% to 70%:
  Additional cached tokens = total_input_qty * (0.70 - X/100)
  Savings = additional_cached_tokens * (regular_rate - cache_read_rate)

3. Batch Inference Opportunities

Batch inference costs 50% less than on-demand for supported models.

Detection: Identify high-volume, non-real-time workloads by looking at:

• Cohere Embed/Rerank usage (typically batch-friendly)
• High token volumes in off-peak hours (requires CloudWatch if available)

Savings estimate: batch_eligible_cost * 0.50

4. Cross-Region Cost Arbitrage

Bedrock pricing varies by region. Compare per-model costs across active regions.

Detection: Group costs by REGION + SERVICE. Flag if same model is used in a more expensive region when a cheaper region is available.

5. Output Token Efficiency

Output tokens cost 5x input tokens. Reducing verbose outputs saves significantly.

Detection: Compare output token cost as % of total. If output > 40% of total token cost, flag for review.

Recommendations:

• Set max_tokens limits on API calls
• Use structured JSON output instead of verbose prose
• Implement response compression for downstream consumers

6. Unused/Low-Value Model Spend

Detect models with minimal usage that may be experimental or abandoned.

Detection: Any model with < $5/month spend across all regions.

Output Format

The script outputs 6 sections. Section 6 ("UNIFIED SAVINGS SUMMARY") is the single source of truth for all savings figures.

Section 6 Output Contract

area    priority    conservative    aggressive
Model Tiering    P0    {opus*0.50}    {opus*0.80}
Cache to 80%    P1    {cache_savings}    {cache_savings}
Batch Inference    P2    {search*0.50}    {search*0.50}
Output Tokens    P2    {output*0.10}    {output*0.20}
TOTAL    -    {sum of above}    {sum of above}
TOTAL_PCT    -    {conservative/total*100}%    {aggressive/total*100}%
TOTAL_SPEND    -    {total_spend}    {total_spend}

Dashboard mapping:

• KPI value: $TOTAL_conservative - $TOTAL_aggressive USD
• KPI trend: +TOTAL_PCT_conservative-TOTAL_PCT_aggressive% of spend
• Table rows: one per area line
• Table summaryRow: TOTAL line

Present as a dashboard artifact with:

• KPI row: Total spend, estimated savings (from TOTAL), savings % (from TOTAL_PCT), top opportunity
• Optimization opportunities table (rows from Section 6 areas)
• Model tiering analysis (Opus vs Sonnet vs Haiku spend breakdown)
• Cache efficiency gauges (hit ratio, reuse ratio vs targets)
• Monthly savings projection chart
• Prioritized recommendations callout

Integration with analyzing-bedrock-costs

This skill builds on data from analyzing-bedrock-costs. The workflow:

1. Run analyzing-bedrock-costs for raw cost/cache data
2. Run this skill's optimize.sh for optimization-specific analysis
3. Use Section 6 output as the single source for all savings in dashboards