ray
40 lines
1.7 KiB
Markdown
40 lines
1.7 KiB
Markdown
# DeepSeek vs Local Ollama — Cost Comparison
|
|
|
|
## Per-Scan Cost (Appointment Screenshot Extraction)
|
|
|
|
| | DeepSeek V4 Flash | Local qwen2.5:14b (Ollama) |
|
|
|---|---|---|
|
|
| OCR step | Free (browser Tesseract.js) | Free (browser Tesseract.js) |
|
|
| LLM input tokens | ~2,000 tokens | ~2,000 tokens |
|
|
| LLM output tokens | ~400 tokens (non-thinking) | ~400 tokens |
|
|
| **Per-scan cost** | **$0.0004** | **$0.0015** (GPU power) |
|
|
| **Scans per dollar** | ~2,500 | ~667 |
|
|
| **Annual (daily)** | ~$0.15 | ~$0.55 |
|
|
|
|
## Token Breakdown
|
|
|
|
Typical appointment screenshot (10 appointments):
|
|
- System prompt: ~1,200 tokens (extraction rules, OCR artifact fixes, VIN recovery)
|
|
- OCR text: ~800 tokens (raw text from 10 appointment rows)
|
|
- Output JSON: ~400 tokens (structured appointments array)
|
|
- Total: ~2,400 tokens
|
|
|
|
## Pricing (DeepSeek V4 Flash, June 2026)
|
|
|
|
| | Per 1M tokens |
|
|
|---|---|
|
|
| Input (cache miss) | $0.14 |
|
|
| Output | $0.28 |
|
|
|
|
Note: Cache hits are rare for one-off scan requests. The system prompt is large (~1,200 tokens) and repeatable across scans, so cached input could drop cost 95% if DeepSeek's KV cache covers it.
|
|
|
|
## Thinking Mode
|
|
|
|
V4 Flash defaults to **thinking mode** (adds ~200 internal reasoning tokens to output). Disabling via `thinking: {type: "disabled"}` keeps costs at the lower rate. The appointment extraction task is straightforward structured parsing — thinking adds cost with no accuracy benefit.
|
|
|
|
## Privacy Trade-off
|
|
|
|
- **DeepSeek**: Customer names, phone numbers, VINs, and vehicle details are sent to DeepSeek's cloud API servers
|
|
- **Ollama**: All data stays on the local GPU (RTX 2080 Ti FE, 11GB VRAM)
|
|
- Cost difference is negligible at ~$0.40/year — the deciding factor is privacy preference
|