Language: English | Español

Cost Analysis and Optimization

Detailed breakdown of LLM normalization costs and strategies to minimize spend

Executive Summary

Production costs (652 leads, 4,280 fields normalized):

Total: $0.043
Per lead: $0.000066 (6.6 cents per 1,000 leads)
Per field: $0.00001 (1 cent per 1,000 fields)

Projected monthly costs (various scales):

Monthly Volume	Total Cost	Per Lead
500 leads	$0.033	$0.000066
1,000 leads	$0.066	$0.000066
5,000 leads	$0.330	$0.000066
10,000 leads	$0.660	$0.000066
50,000 leads	$3.300	$0.000066

Cost is linear and predictable - $0.066 per 1,000 leads regardless of scale.

Cost Breakdown

AWS Bedrock: Claude 3 Haiku Pricing

Pricing (as of January 2026):

Input tokens: $0.00025 per 1,000 tokens ($0.25 per 1M tokens)
Output tokens: $0.00125 per 1,000 tokens ($1.25 per 1M tokens)

Per-batch cost (10 leads, 7 fields each):

Component	Tokens	Cost
Prompt (fixed overhead)	~800	$0.0002
Lead data (10 leads)	~500	$0.00013
Total input	~1,300	$0.00033
Output (JSON response)	~80	$0.0001
Total per batch	~1,380	$0.00043

Cost per lead: $0.00043 / 10 = $0.000043

Actual production costs (652 leads):

Batches: 65 batches (652 / 10)
Total input tokens: ~84,500 tokens
Total output tokens: ~5,200 tokens
Total cost: $0.028 (Bedrock) + $0.015 (Lambda/DynamoDB) = $0.043

Lambda Costs

Configuration:

Memory: 512 MB
Duration: ~3 minutes for 50 leads
Executions: 1 per day (scheduled)

Cost calculation:

Monthly executions: 30
Duration per execution: 180s (3 min)
GB-seconds: (512 MB / 1024) × 180s × 30 = 2,700 GB-s

Lambda pricing:
- First 400,000 GB-s/month: FREE (well within free tier)
- Requests: 30/month (FREE - first 1M requests free)

Lambda monthly cost: $0.00

Lambda is effectively free for this workload due to AWS Free Tier.

DynamoDB Costs

Operations per normalization:

Reads: 1 × GetItem (config) + 1 × Scan (leads) = ~100 read units
Writes: 50 × UpdateItem (normalized leads) = 50 write units

Cost calculation (pay-per-request):

Monthly reads: 100 reads × 30 days = 3,000 reads
Monthly writes: 50 writes × 30 days = 1,500 writes

DynamoDB pricing:
- Reads: $0.25 per 1M reads → $0.00075
- Writes: $1.25 per 1M writes → $0.001875

DynamoDB monthly cost: $0.0026 (~$0.003)

EventBridge Costs

Invocations: 1 per day × 30 days = 30 invocations/month

Cost: First 1 million events are FREE.

EventBridge monthly cost: $0.00

Total Monthly Cost (1,000 Leads/Month)

Service	Cost
AWS Bedrock (Claude 3 Haiku)	$0.066
Lambda	$0.000 (Free Tier)
DynamoDB	$0.003
EventBridge	$0.000 (Free Tier)
Total	$0.069

Cost per lead: $0.000069 (~7 cents per 1,000 leads)

Cost Comparison: Alternatives

1. Manual Data Entry

Scenario: Hire a data entry clerk at $15/hour to normalize 1,000 leads.

Assumptions:

5 minutes per lead (check 7 fields, standardize formats)
Hourly rate: $15

Cost:

Time: 1,000 leads × 5 min = 5,000 min = 83.3 hours
Cost: 83.3 hours × $15/hour = $1,250

Comparison: This pattern is 18,000x cheaper than manual entry.

2. Custom Rule-Based ETL

Scenario: Build a custom Python script with regex patterns and lookup tables.

Costs:

Development time: 2 weeks (80 hours) × $75/hour = $6,000
Maintenance: 4 hours/month × $75/hour = $300/month
Runtime: Lambda free tier, DynamoDB minimal

First-year cost: $6,000 + ($300 × 12) = $9,600

Comparison: This pattern pays for itself in < 1 month if:

You value developer time
You need flexibility (LLM adapts to new patterns without code changes)
You don't want to maintain hundreds of regex rules

3. Claude 3.5 Sonnet (Higher-End LLM)

Scenario: Use Sonnet instead of Haiku for normalization.

Sonnet pricing:

Input: $0.003 per 1,000 tokens (12x more expensive)
Output: $0.015 per 1,000 tokens (12x more expensive)

Cost per 1,000 leads:

Haiku: $0.066
Sonnet: $0.792

Difference: $0.726 more per 1,000 leads (12x increase)

When to use Sonnet:

Complex reasoning required (not the case for normalization)
Quality improvement justifies 12x cost (rare for structured data)

Recommendation: Stick with Haiku unless quality metrics drop below 95%.

4. GPT-4o-mini (OpenAI Alternative)

Scenario: Use OpenAI GPT-4o-mini instead of Claude Haiku.

GPT-4o-mini pricing:

Input: $0.00015 per 1,000 tokens (40% cheaper than Haiku)
Output: $0.0006 per 1,000 tokens (52% cheaper than Haiku)

Cost per 1,000 leads:

Haiku: $0.066
GPT-4o-mini: $0.040

Difference: $0.026 savings per 1,000 leads (40% cheaper)

Considerations:

Pros: Cheaper, widely available
Cons: Requires OpenAI API (separate billing), different API format, may require prompt tuning

Recommendation: If already using OpenAI API, consider GPT-4o-mini. Otherwise, Haiku's integration with AWS (same billing, IAM auth) outweighs marginal cost savings.

Cost Optimization Strategies

1. Increase Batch Size

Current: 10 leads per API call

Optimization: 20 leads per API call

Impact:

Current: 1,000 leads → 100 batches × $0.00043 = $0.043
Optimized: 1,000 leads → 50 batches × $0.00055 = $0.028  (35% savings)

Why it works: Prompt overhead is amortized over more leads.

Trade-offs:

⚠️ Higher memory usage (512 MB → 768 MB might be needed)
⚠️ Longer timeout needed (300s might not suffice for 20-lead batches)
✅ Reduced API calls (better for rate limiting)

Recommendation: Test with 15-20 leads per batch, monitor Lambda memory/timeout.

2. Reduce Fields Normalized

Current: 7 fields per lead

Optimization: Normalize only high-priority fields (4 fields)

Impact:

Current: 7 fields × 1,000 leads = 7,000 fields → $0.066
Optimized: 4 fields × 1,000 leads = 4,000 fields → $0.038  (42% savings)

Trade-offs:

⚠️ Less comprehensive normalization
✅ Lower costs
✅ Faster processing

Recommendation: Only normalize fields that impact reporting/analytics.

Priority matrix:

Priority	Fields	Why
High	ciudad, nivelEducativo, ocupacionActual	High variance, impact analytics
Medium	empresa, direccion	Moderate variance, nice to have
Low	nombres, apellidos	Low variance, mostly formatted

3. Lengthen Normalization TTL

Current: Re-normalize every 7 days

Optimization: Re-normalize every 30 days

Impact:

Current: 1,000 leads normalized 4 times/month = 4,000 normalizations/month
Optimized: 1,000 leads normalized 1 time/month = 1,000 normalizations/month

Savings: 75% reduction in re-normalization costs

Trade-offs:

⚠️ Stale normalized data if prompts/models change
✅ 4x cost reduction for re-normalizations

Recommendation: 7-day TTL for active data, 30-day TTL for archived data.

4. Use Caching for Repeated Values

Scenario: Many leads share the same city/company.

Optimization: Cache normalized values in memory.

const cache = new Map();

function normalizeCityWithCache(city) {
  const key = city.toLowerCase();

  if (cache.has(key)) {
    return cache.get(key);  // Skip LLM call
  }

  const normalized = await normalizeLead({ ciudad: city });
  cache.set(key, normalized.ciudad);
  return normalized.ciudad;
}

Impact:

Example: 1,000 leads with only 50 unique cities
Without cache: 1,000 API calls
With cache: 50 API calls (95% reduction)

Savings: $0.066 → $0.003 (95% cheaper)

Trade-offs:

⚠️ Requires cache invalidation strategy
⚠️ Doesn't work well for unique fields (names, addresses)
✅ Massive savings for fields with low cardinality

Recommendation: Implement for ciudad, empresa, nivelEducativo (low cardinality fields).

5. Skip Already-Normalized Leads

Current: Always re-normalize if TTL expired

Optimization: Hash normalized data, skip if unchanged

async function normalizeLead(lead) {
  const fieldsData = extractFields(lead);
  const currentHash = hashFields(fieldsData);

  if (lead.normalizedHash === currentHash) {
    console.log('Data unchanged, skipping normalization');
    return { normalized: false, reason: 'Unchanged' };
  }

  // Proceed with normalization...
}

Impact:

Scenario: 30% of leads haven't changed since last normalization
Without optimization: 1,000 normalizations
With optimization: 700 normalizations

Savings: 30% cost reduction

Recommendation: Implement if re-normalization costs become significant.

Scaling Cost Projections

Small Scale (1,000 Leads/Month)

Service	Cost
Bedrock	$0.066
Lambda	$0.000
DynamoDB	$0.003
Total	$0.069

Cost per lead: $0.000069

Medium Scale (10,000 Leads/Month)

Service	Cost
Bedrock	$0.660
Lambda	$0.005
DynamoDB	$0.030
Total	$0.695

Cost per lead: $0.000069 (same - linear scaling)

Large Scale (100,000 Leads/Month)

Service	Cost
Bedrock	$6.60
Lambda	$0.50
DynamoDB	$3.00
Total	$10.10

Cost per lead: $0.000101 (marginal increase due to Lambda/DynamoDB)

Very Large Scale (1,000,000 Leads/Month)

Service	Cost
Bedrock	$66.00
Lambda	$5.00
DynamoDB	$30.00
Total	$101.00

Cost per lead: $0.000101

At this scale, consider:

Reserved capacity for Lambda/DynamoDB (20-30% savings)
Caching layer (Redis/ElastiCache) to reduce redundant normalizations
Fine-tuning a smaller model (Llama 3) for on-premise inference

ROI Analysis

Scenario: Educational Institution (This Project)

Context: 100 students/month enroll, 652 leads collected.

Costs:

Normalization: $0.043 (one-time for 652 leads)
Infrastructure: $0.003/month (DynamoDB)

Benefits:

Reporting accuracy: Clean data enables accurate analytics
Time savings: 5 hours/month saved on manual data cleaning (5 × $50/hour = $250/month)
AI evaluation: Improved prompt quality for downstream AI candidate evaluation

ROI: (250 - 0.043 - 0.003) / 0.046 = 5,400% ROI (54x return)

Scenario: SaaS Platform (10,000 Users/Month)

Context: B2B SaaS with user-submitted company/industry data.

Costs:

Normalization: $0.69/month (10,000 users × 3 fields)
Developer time saved: 20 hours/month × $100/hour = $2,000/month

ROI: (2,000 - 0.69) / 0.69 = 289,800% ROI (2,898x return)

Scenario: E-commerce (100,000 Products/Month)

Context: Marketplace with seller-submitted product descriptions.

Costs:

Normalization: $10.10/month (100,000 products × 5 fields)
Manual QA reduction: 50 hours/month × $75/hour = $3,750/month

ROI: (3,750 - 10.10) / 10.10 = 37,000% ROI (370x return)

Budget Planning

Monthly Budget Template

Base costs (pay-per-request):
  Bedrock: $0.066 per 1,000 leads
  Lambda: $0.000 (Free Tier up to 400,000 GB-s)
  DynamoDB: $0.003 per 1,000 leads
  EventBridge: $0.000 (Free Tier up to 1M events)

Total: $0.069 per 1,000 leads

Scaling factors:
  × Number of thousands of leads
  × Average fields per lead / 7 (baseline)
  × Re-normalization frequency / 30 days (baseline)

Example (5,000 leads/month, 10 fields, weekly re-normalization):
  Base: $0.069 × 5 = $0.345
  Field adjustment: × (10 / 7) = $0.493
  Frequency adjustment: × (30 / 7) = $2.11/month

Cost Alerts

Recommended CloudWatch Billing Alarms:

Volume	Monthly Budget	Alert Threshold
1,000 leads	$0.07	$0.10 (safety margin)
10,000 leads	$0.70	$1.00
100,000 leads	$10.00	$15.00

SAM template:

CostAlarm:
  Type: AWS::CloudWatch::Alarm
  Properties:
    AlarmName: normalization-cost-spike
    Namespace: AWS/Billing
    MetricName: EstimatedCharges
    Dimensions:
      - Name: ServiceName
        Value: AmazonBedrock
    Statistic: Maximum
    Period: 86400  # 24 hours
    EvaluationPeriods: 1
    Threshold: 1.00  # $1/day = ~$30/month
    ComparisonOperator: GreaterThanThreshold
    AlarmActions:
      - !Ref AlertTopic

Conclusion

This pattern is extremely cost-effective:

$0.066 per 1,000 leads (7 fields each)
Linear scaling (predictable costs)
18,000x cheaper than manual data entry
Pays for itself in < 1 month vs custom development

Key cost drivers:

Bedrock API calls (94% of total cost) → optimize by batching
Number of fields (linear) → prioritize high-value fields
Re-normalization frequency (multiplicative) → adjust TTL based on data volatility

Optimization priorities:

Increase batch size (35% savings, minimal risk)
Implement caching (up to 95% savings for low-cardinality fields)
Reduce re-normalization frequency (75% savings for stable data)

Bottom line: At $0.000066 per lead, this pattern is a no-brainer for any system with >100 leads/month.

Next Steps

README.md: Pattern overview and quick start
IMPLEMENTATION.md: Step-by-step setup guide
STATISTICAL-VALIDATION.md: Quality measurement
LESSONS-LEARNED.md: Production insights

Last Updated: January 24, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Cost Analysis and Optimization

Executive Summary

Cost Breakdown

AWS Bedrock: Claude 3 Haiku Pricing

Lambda Costs

DynamoDB Costs

EventBridge Costs

Total Monthly Cost (1,000 Leads/Month)

Cost Comparison: Alternatives

1. Manual Data Entry

2. Custom Rule-Based ETL

3. Claude 3.5 Sonnet (Higher-End LLM)

4. GPT-4o-mini (OpenAI Alternative)

Cost Optimization Strategies

1. Increase Batch Size

2. Reduce Fields Normalized

3. Lengthen Normalization TTL

4. Use Caching for Repeated Values

5. Skip Already-Normalized Leads

Scaling Cost Projections

Small Scale (1,000 Leads/Month)

Medium Scale (10,000 Leads/Month)

Large Scale (100,000 Leads/Month)

Very Large Scale (1,000,000 Leads/Month)

ROI Analysis

Scenario: Educational Institution (This Project)

Scenario: SaaS Platform (10,000 Users/Month)

Scenario: E-commerce (100,000 Products/Month)

Budget Planning

Monthly Budget Template

Cost Alerts

Conclusion

Next Steps

FilesExpand file tree

COST-ANALYSIS.md

Latest commit

History

COST-ANALYSIS.md

File metadata and controls

Cost Analysis and Optimization

Executive Summary

Cost Breakdown

AWS Bedrock: Claude 3 Haiku Pricing

Lambda Costs

DynamoDB Costs

EventBridge Costs

Total Monthly Cost (1,000 Leads/Month)

Cost Comparison: Alternatives

1. Manual Data Entry

2. Custom Rule-Based ETL

3. Claude 3.5 Sonnet (Higher-End LLM)

4. GPT-4o-mini (OpenAI Alternative)

Cost Optimization Strategies

1. Increase Batch Size

2. Reduce Fields Normalized

3. Lengthen Normalization TTL

4. Use Caching for Repeated Values

5. Skip Already-Normalized Leads

Scaling Cost Projections

Small Scale (1,000 Leads/Month)

Medium Scale (10,000 Leads/Month)

Large Scale (100,000 Leads/Month)

Very Large Scale (1,000,000 Leads/Month)

ROI Analysis

Scenario: Educational Institution (This Project)

Scenario: SaaS Platform (10,000 Users/Month)

Scenario: E-commerce (100,000 Products/Month)

Budget Planning

Monthly Budget Template

Cost Alerts

Conclusion

Next Steps