AI Browser Automation Cost Analysis (Part 1): Tokens, Proxy, and Compute

A Typical Task Cost Breakdown

10,000 pages/day, 3-5 fields per page. Costs come from three dimensions:

Dimension 1: LLM Inference Cost

Daily cost for 10,000 tasks: $90 (Gemini Flash) to $1,800 (Claude Sonnet).

With A11y tree optimization, token consumption drops 90%: $9 to $180.

Dimension 2: Proxy Traffic Cost

Proxy is usually 5-15% of total cost. Usually not worth optimizing.

Dimension 3: Compute Cost

Compute is 10-20% of total.

Total Cost Structure

LLM inference: 70-85%  → optimize here
Proxy traffic: 5-15%   → not worth much effort
Compute: 10-20%        → optimize after LLM

Five Effective Optimizations

1. DOM Cache

class DOMCache:
    def __init__(self, ttl_seconds=300):
        self.cache = {}
        self.ttl = ttl_seconds
 
    async def get_dom(self, url, page):
        if url in self.cache:
            entry = self.cache[url]
            if time.time() - entry["time"] < self.ttl:
                return entry["dom"]
        dom = await extract_dom(page)
        self.cache[url] = {"dom": dom, "time": time.time()}
        return dom

2. Reduce Screenshot Frequency

Screenshots are expensive. Base64-encoded at ~500KB, they consume many tokens in vision models. Only capture for debugging and critical steps.

3. Reuse Sessions

agent-browser auth save ./session.json
agent-browser auth login ./session.json
agent-browser open https://example.16yun.cn

4. Choose the Right Model

Don't use the strongest model for everything. Simple tasks → Gemini Flash or Haiku. Complex → Sonnet or GPT-4o.

5. Don't Over-Proxy

If the target site doesn't have anti-bot, datacenter proxy is sufficient. Only upgrade when detected.

What Not to Optimize

Summary

LLM inference is 70-85% of total cost. Optimization priorities:

1. Token consumption (A11y tree, DOM cache, fewer screenshots)
2. Model selection (match model to task complexity)
3. Session reuse (reduce login token waste)
4. Proxy and compute (last priority)