← Back
Economics
Cost breakdown per extraction and infrastructure costs at scale.
Per-Extraction Cost
Classification (Haiku, 1 page)$0.003
Client extraction (Haiku, 2 pages)$0.005
Row extraction (Sonnet, 2-4 pages avg)$0.025
Email context (Haiku, if email)$0.002
Snake API matching (batch)$0.000
Normalization (deterministic)$0.000
Total per extraction~$0.033
Model Pricing (Bedrock)
| Model | Input | Output | Use |
|---|
| Claude Haiku 4.5 | $0.80/M tokens | $4.00/M tokens | Classify, client, email context |
| Claude Sonnet 4.6 | $3.00/M tokens | $15.00/M tokens | Row extraction (vision) |
| Snake API | Free (internal) | — | Article matching |
Infrastructure
| Resource | Spec | Monthly Cost |
|---|
| EC2 | t3.medium (2 vCPU, 4GB) | $30 |
| EBS | 30GB gp3 | $2.40 |
| Data transfer | ~50GB out | $4.50 |
| SSL | Let's Encrypt (free) | $0 |
| Total infra | | ~$37/mo |
At Scale
| Volume | LLM Cost | Infra Cost | Total | Per Extraction |
|---|
| 100/day | $99/mo | $37/mo | $136/mo | $0.045 |
| 500/day | $495/mo | $37/mo | $532/mo | $0.035 |
| 2000/day | $1,980/mo | $74/mo (2x) | $2,054/mo | $0.034 |
| 10,000/day | $9,900/mo | $148/mo (4x) | $10,048/mo | $0.033 |
vs. Manual Processing
Human operator (avg 5 min/doc, €25/hr)€2.08/doc
Glass Pipeline (automated)€0.03/doc
Cost reduction98.5%
At 500 docs/day: manual = €31,200/mo vs. automated = €532/mo. Saves €30,668/mo.
Latency Budget
PDF Parse: 50-200ms (local, PyMuPDF)
Classification: 800-1200ms (Haiku VLM)
Client extract: 800-1500ms (Haiku VLM, parallel with extraction)
Row extraction: 8-25s (Sonnet VLM, depends on pages)
Normalization: <1ms (deterministic)
Snake matching: 50-200ms (batch HTTP)
─────────────────────────────────────
Total: 10-28s (median ~15s)