SOM Cost Analysis
How much does SOM compression save on LLM input costs? We tested 49 real-world websites and measured the token reduction from raw HTML to Plasmate SOM output.
16.6x
Overall Compression
10.5x
Median Compression
94%
Token Cost Savings
49
URLs Tested
Range: 0.7x (example.com, tiny page) to 116.9x (cloud.google.com)
Cost per 1,000 Page Loads (49 URLs)
What it costs to send the content of all 49 test URLs to an LLM, repeated 1,000 times:
| Model | HTML Cost | SOM Cost | Savings |
|---|---|---|---|
| GPT-4 ($10/M) | $50,397 | $3,042 | $47,355 (94%) |
| GPT-4o ($2.50/M) | $12,599 | $761 | $11,839 (94%) |
| Claude Sonnet ($3/M) | $15,119 | $913 | $14,207 (94%) |
At Scale: 1M Pages/Month
Monthly cost for processing one million pages at the average page size from this benchmark:
| Model | HTML/month | SOM/month | Monthly Savings |
|---|---|---|---|
| GPT-4 | $1,028 | $62 | $966 |
| GPT-4o | $257 | $16 | $241 |
| Sonnet | $308 | $19 | $290 |
Top 15 Compression Results
cloud.google.com
116.9x
linear.app
105.3x
reddit.com
103.8x
figma.com
63.6x
tailwindcss.com
53.7x
nodejs.org
42.7x
vercel.com
35.7x
wired.com
35.0x
aws.amazon.com
31.4x
mongodb.com
29.2x
typescriptlang.org
27.7x
theguardian.com
20.3x
azure.microsoft.com
15.7x
nytimes.com
15.3x
stripe.com/docs
14.6x
Token Counts
| Site | HTML Tokens | SOM Tokens | Ratio |
|---|---|---|---|
| cloud.google.com | 464,616 | 3,973 | 116.9x |
| linear.app | 562,922 | 5,347 | 105.3x |
| reddit.com | 121,027 | 1,166 | 103.8x |
| figma.com | 369,124 | 5,806 | 63.6x |
| tailwindcss.com | 233,071 | 4,338 | 53.7x |
| nodejs.org | 115,771 | 2,709 | 42.7x |
| vercel.com | 198,761 | 5,565 | 35.7x |
| wired.com | 350,411 | 10,006 | 35.0x |
| aws.amazon.com | 81,318 | 2,588 | 31.4x |
| mongodb.com | 232,141 | 7,944 | 29.2x |
| typescriptlang.org | 64,899 | 2,345 | 27.7x |
| theguardian.com | 319,879 | 15,753 | 20.3x |
| azure.microsoft.com | 134,654 | 8,555 | 15.7x |
| nytimes.com | 48,042 | 3,133 | 15.3x |
| stripe.com/docs | 48,378 | 3,305 | 14.6x |
Methodology
- Each URL fetched once via
plasmate fetch <url> - HTML token count: raw HTML bytes / 4 (standard byte-to-token estimate)
- SOM token count: SOM JSON output bytes / 4
- Compression ratio: HTML tokens / SOM tokens
- Cost calculated at published model input pricing
- No caching, no warm-up, single fetch per URL
- 2 URLs failed (w3.org, dev.to) and were excluded
Reproduce
cargo install plasmate
git clone https://github.com/plasmate-labs/plasmate
cd plasmate
./benchmarks/run-cost-analysis.sh