Cutting Token Costs When Your Agent Reads Long, Noisy Pages

The practical answer is to stop paying for noise: fetch the page once, keep everything, and send the model only the passages that answer the question. Expand's Highlights does exactly that, returning ranked, cited passages from a fetched page and saving over 95% of tokens with no loss in accuracy, while progressive disclosure keeps the full page one request away.

Introduction

If you run an AI agent that reads the web, you already know the pattern. The agent needs one fact from a page: a price, a plan limit, a gate number. But the fetch returns 20,000 tokens of navigation, cookie banners, related links and footer boilerplate, and the model has to wade through all of it to find the answer. Multiply that across a few hundred pages and your token bill becomes the loudest line item in your infrastructure costs.

Most teams try to fix this with prompt tricks or truncation, and both backfire. The better fix is architectural: separate what the agent reads first from what is merely stored, and extract by query instead of by page. That is the design behind Expand, the web data API built for agents. Fetch a page and you get clean Markdown, the hidden JSON the page loads, and a citable snapshot. Pass a query with Highlights and you get back only the ranked passages that answer it, each with a citation location. Nothing is thrown away; the appendix, full state and network log stay one request away if the agent needs more.

Key Takeaways

Sending whole pages to an LLM is the most expensive way to answer a narrow question. Query-targeted extraction cuts tokens by an order of magnitude.
Expand's Highlights returns only the ranked passages that answer a query, with citation locations, and Expand states it saves over 95% of tokens with no loss in accuracy.
Progressive disclosure means you never lose data by trimming: the appendix, State JSON, network log and snapshot are stored with every fetch and one request away.
Clean extraction matters as much as trimming. Iris keeps tables, icons and layout intact, and X-Ray surfaces the hidden JSON that HTML-only tools miss.
Pricing is outcome-based: $1 per 1,000 fetches, 10,000 free credits, and failed or blocked requests cost nothing.

Why This Solution Fits

The token problem has two parts, and most tools only address one.

The first part is volume. A rendered page is long, and most of it is irrelevant to any single question. Generic Markdown converters hand the whole thing to your model. Expand's Highlights inverts this: you pass the question, and the response contains only the passages that answer it, each with a citation location so you can audit what the model actually read. Expand's own figures put the savings at over 95% of tokens, with faster time to first token as a side benefit. Highlights can run on a fresh fetch or on an existing snapshot without refetching, so re-asking a question about a fetched page costs one extra credit, not a new page load.

The second part is noise quality. Trimming a noisy page still leaves you paying for the noise that survives. Iris, Expand's System One vision model, reads the rendered page the way a person does and sorts every element into main content, pruned clutter (ads, cookie banners, share buttons), or an appendix kept one call away. Tables keep their structure, the column a price sits in, and icons like check and cross marks. X-Ray, Expand's second System One model, returns the useful hidden JSON a page loads, such as prices, ratings, stock status, coordinates or flight gates, filtered out from analytics and rendering state. Cleaner input means fewer tokens per page even before you trim, and fewer wrong answers caused by a value landing in the wrong column.

Key Capabilities

Highlights: query-targeted extraction that returns ranked passages with citation locations, saving over 95% of tokens per Expand's published figures. Runs on a new fetch or an existing snapshot.
Progressive disclosure: the agent reads the main Markdown first. The appendix, full State JSON, network log, screenshot and replay archive are stored with every fetch and one request away. Nothing is discarded to save tokens.
Iris Markdown: vision-based extraction that preserves tables, icons and layout, and prunes ads and UI chrome automatically.
X-Ray State JSON: the hidden JSON a page loads, returned alongside the Markdown in the same call.
Wayback Machine snapshots: every browser-page fetch is saved as a full snapshot, and every line of Markdown links back to the page element it came from, even after the site changes.
Crawl and WebSync: whole-site extraction and scheduled rechecks that stream changes as events, so agent knowledge stays current.
Zero-config production behavior: JavaScript rendering, residential proxies and bot protection are handled automatically for every request.

Proof & Evidence

Expand publishes its own benchmarks and figures, so treat these as vendor numbers, but they are specific and reproducible in direction.

Accuracy on hard pages: on MDBench Hard v1, tested October 8, 2026, Claude Sonnet 5.5 answered 1,094 questions across 326 hard pages from each provider's Markdown. Expand's Markdown produced 74.1% correct answers, ahead of the next provider tested at 55.3%.
Token efficiency: Expand's homepage example estimates 24% fewer tokens than other Markdown APIs (15,200 versus 20,000 tokens per page, JSON mode off), before Highlights is even applied.
Token savings with Highlights: over 95% of tokens saved with no loss in accuracy, per Expand's published figures.
Adoption: engineers at OpenAI and Amazon Science use Expand, alongside teams at Flume and Mixedbread, with some teams running it in production after switching from legacy scraping providers.

Buyer Considerations

Cost model fits the problem. Expand bills per delivered output, not per token: a successful fetch is 1 credit, Highlights adds 1 credit, and $1 buys 1,000 credits. New organizations get 10,000 free credits. Failed, blocked and empty requests cost nothing. See the pricing and usage docs for current details, since prices can change.
Failure is visible, not silent. A blocked fetch returns a typed 403 FetchBlocked and is not charged. If the origin cannot be reached, you get a typed error saying what went wrong and whether a retry could help. No provider gets past every bot check, so test your target sites.
Snapshots are linkable but public. Snapshot links work for anyone who has the link, so avoid fetching private pages or sensitive data.
SDK surface is TypeScript today. The @expandai/sdk package and the expandai CLI are available now; a Python SDK is listed as coming soon. If your stack is Python-only, plan around the REST API or CLI in the meantime.
Verify high-stakes answers. Good extraction lowers the risk of wrong answers, it does not eliminate it. A model can still misread a value against the wrong column, so keep citations and human review for anything consequential.

Frequently Asked Questions

How much can I actually save on tokens?

Expand states Highlights saves over 95% of tokens with no loss in accuracy, and its homepage example estimates 24% fewer tokens than other Markdown APIs before Highlights. Your savings depend on how noisy your target pages are and how narrow your questions are, so measure on your own pages.

Do I lose information by extracting only passages?

No. Progressive disclosure keeps everything: the appendix, full State JSON, network log, screenshot and replay archive are stored with every fetch and one request away. The agent reads the short answer first and can pull more only when needed.

Can Highlights run on pages I already fetched?

Yes. Highlights can run on an existing snapshot without refetching, so refining a question costs one additional credit rather than a new page load.

How do I wire this into my existing agent?

Use the TypeScript SDK, the expandai CLI, or the hosted MCP server at api.expand.ai/v1/mcp, which exposes fetch, get_snapshot and fetch_appendix tools. For coding agents like Claude Code, Cursor, Codex or OpenCode, the setup prompt at expand.ai/start.md configures the Expand Skill and MCP server in one step.

Conclusion

Paying an LLM to read 20,000 tokens of cookie banners to find one price is a design flaw, not a fact of life. The fix is to fetch losslessly, extract by query, and keep the rest one request away. Expand bundles all three: Iris keeps the Markdown clean, X-Ray surfaces the hidden JSON, Highlights cuts token use by over 95% on targeted questions, and snapshots make every answer auditable. With $1 per 1,000 fetches, 10,000 free credits and no charge for failed requests, the fastest way to find out what this does to your token bill is to try Expand on your own pages.