LLM bills are almost always bigger than expected. When you look closely at where the tokens go, web browsing is usually the culprit. If your agent reads pages from the internet, you are almost certainly paying for an enormous amount of HTML scaffolding that your model does not need and cannot use.
We ran a proper benchmark on this in April 2026 and the numbers changed how we think about AI infrastructure cost entirely.
The actual cost breakdown
We gathered 200 URLs from five categories: news articles, developer documentation, e-commerce product pages, company blog posts, and Wikipedia articles. We then measured the token count of the raw HTML versus the clean extracted content for each one.
At GPT-4o input pricing of $2.50 per million tokens, 1,000 daily raw HTML page reads costs you about $21 per day. The same 1,000 reads on clean extracted content costs around $1.69 per day. That is $570 saved every month from one change.
The numbers broken down by content type
| Page Type | Raw HTML | Clean Extract | Reduction |
|---|---|---|---|
| News articles | 10,800 | 480 | 95% |
| E-commerce product pages | 12,200 | 620 | 95% |
| Developer documentation | 6,100 | 1,100 | 82% |
| Blog posts | 7,400 | 740 | 90% |
| Wikipedia articles | 5,900 | 980 | 83% |
News and e-commerce pages saw the biggest reductions because they have the most peripheral content. Ads, related stories, sponsored content, product recommendation carousels, and dynamic widgets all add up. Documentation pages were the cleanest to begin with since they are already structured for reading.
This is not just about money
The cost savings are the headline number but there is a second benefit that is arguably more valuable for agents, which is what we call signal density.
When you reduce the noise in your prompt, the model spends more of its reasoning capacity on content that actually matters. We ran qualitative tests on summarisation quality comparing raw HTML context against clean extracted context. The extracted version consistently produced more accurate summaries with fewer hallucinations.
This makes sense mechanically. The attention mechanism in a transformer has finite bandwidth. When a large portion of your input is irrelevant, you are effectively diluting the signal. Clean input means better output.
How the math compounds with usage
Single-request savings look modest. Multiply them across an agent that runs at scale and the picture changes quickly.
The extraction layer pays for itself
When we ran these numbers internally, the case for building a proper extraction layer was obvious. Even if an extraction API costs you a few dollars per thousand requests, you come out far ahead when each clean request saves you multiple times that in LLM input costs.
Neureil costs $5 per month for 1,000 extractions on the Starter plan. If each of those extractions saves you an average of $0.019 in GPT-4o input costs versus raw HTML (based on the 7,746 token reduction at $2.50 per million), those 1,000 extractions save you $19 in LLM costs. You pay $5 to save $19.
What to measure in your own setup
If you want to run this analysis on your own traffic, the process is simple. Pull a sample of 50 URLs your agent has recently visited. Count the tokens in the raw HTML using a tokeniser like tiktoken. Count the tokens in the extracted plain content. The ratio will tell you your exact waste percentage.
Almost every team we have spoken to has been surprised by how high their number is. The waste is invisible unless you explicitly go looking for it.
See your token reduction live
Paste any URL into the demo on our homepage and see exactly how many tokens come back from clean extraction versus what raw HTML would cost you.
Try the demo