AI Browser Automation Cost Analysis (Part 2): Proxy and Compute Optimization
Tokens dominate, but proxy and compute also have room for optimization. Right-sized proxy selection, concurrency strategy, and instance type matching.
16Yun Engineering TeamMay 2, 20261 min read
Proxy Optimization
Proxy cost is small, but wrong proxy selection increases failure rates, which increases token costs (retries need more LLM calls).
Proxy Selection Decision Tree
Target site has anti-bot?
├── No → Datacenter proxy (cheapest)
├── Yes, mild → Crawler Proxy (tunnel)
└── Yes, strict (Turnstile) → Residential/Dedicated + GeoIPInstance Type Selection
Browsers are memory-intensive, not CPU-intensive:
| Cloud | Instance | vCPU | RAM | Monthly (on-demand) |
|---|---|---|---|---|
| AWS | r6i.large | 2 | 16GB | ~$70 |
| AWS | r6i.xlarge | 4 | 32GB | ~$140 |
| GCP | n2-highmem-2 | 2 | 16GB | ~$60 |
Pod Density
A 16GB instance can run approximately:
- 4-6 simple page browser instances
- 2-3 complex SPA instances
- 1-2 multi-tab instances
Beyond this, GC and OOM increase failure rates, raising total cost.
Cost Curve at Scale
| Daily Tasks | Token (Gemini Flash) | Proxy | Compute | Total |
|---|---|---|---|---|
| 1,000 | $9 | $0.10 | $1.50 | ~$11 |
| 10,000 | $90 | $1.00 | $15 | ~$106 |
| 100,000 | $900 | $10 | $150 | ~$1,060 |
| 1,000,000 | $9,000 | $100 | $1,500 | ~$10,600 |
Token cost is always dominant and grows linearly. Proxy and compute can be amortized at scale. Tokens cannot.
Summary
- Don't over-proxy — match proxy type to target site protection level
- Match instance type — use high-memory-ratio instances
- Evaluate warm pools before using — not needed for all scenarios
- No scale economy on tokens — token cost is linear
Need an enterprise proxy plan?
We can tailor architecture to your target domains, concurrency, and reliability goals.