docs: add Costs section to README with token usage and cost breakdown

Adds a new ## Costs section summarizing observed LLM costs based on
maintainer testing (~82 requests, 246k tokens). Includes a stats table
(avg, 95th percentile, peak), a token breakdown table, and a key insight
explaining that output tokens drive ~50% of cost despite being only 17%
of volume. Includes a disclaimer that figures are based on personal testing.

https://claude.ai/code/session_01Raf7fKQax1GZEFoYUsGyxD

Author: claude

README.md

@@ -488,6 +488,33 @@ The PR comment is always posted — with a GIF if recording succeeded, or screen
 
 ---
 
+## Costs
+
+> **Disclaimer:** The figures below are based on the maintainer's own testing and may not reflect your usage patterns. Token consumption varies significantly by diff size, app complexity, and retry count. Use these numbers as a rough reference only — your costs may be higher or lower.
+
+Each GitGlimpse run makes one or more LLM calls to generate the Playwright interaction script. The table below summarizes observed costs using `claude-sonnet-4-6` with default settings.
+
+| Metric | Value |
+|---|---|
+| Average cost per request | ~$0.015 (1.5¢) |
+| 95th percentile | ~$0.035 |
+| Peak (max observed) | ~$0.047 |
+| Requests per USD | ~66 |
+| Sample size | ~82 requests / 246k tokens |
+
+**Token breakdown (per request, approximate):**
+
+| Token type | Volume | Share of tokens | Share of cost |
+|---|---|---|---|
+| Input | ~205k total | ~83% | ~50% |
+| Output | ~42k total | ~17% | ~50% |
+
+**Key insight:** Output tokens drive roughly **50% of the cost** despite representing only **17% of the token volume** — a direct consequence of output tokens being priced ~5× higher than input tokens in most Claude pricing tiers. Reducing script verbosity or capping `maxDuration` can lower output token counts and trim costs noticeably.
+
+For high-frequency teams, use `trigger.mode: 'smart'` or `'on-demand'` to avoid running the LLM on every push.
+
+---
+
 ## Development
 
 ```bash