Expand Fetch: An Agent Web-Reading API Claims 74% on a Hard Benchmark

Agents answer wrongly about web pages more often than people expect, and the cause is often the format they are given. Expand Fetch, launched on October 14, 2026 by Tim Suchanek's company Expand, is a web-reading API built on that diagnosis. Its announcement and homepage claim that agents using it answered 74.1 percent of hard questions correctly, against 55.3 percent for Exa and 27.2 percent for Firecrawl. The benchmark was built and run by Expand, so this is a claimed-versus-verified post.

If you build agents that browse, scrape or research, the failure mode Expand describes is real and worth designing around, whichever tool you use. For how agents get into trouble with web content more broadly, see our guides on bot traffic hitting small sites and indirect prompt injection.

Quick reference

The problem Expand is pointing at

The homepage example is simple. A pricing page marks Business with a cross and Enterprise with a tick next to "Uptime SLA". A converter that drops icons produces two plans that both list the feature, and the agent answers "both". Keep the icons as a table and the agent answers "only Enterprise". The announcement says markdown APIs "drop icons, break tables, and miss hidden JSON", and that you cannot tell whether the page failed to load or the conversion dropped something.

That is a legitimate engineering point. Icons, colour coding and layout carry meaning that plain text conversion loses, and pages often embed richer data in hidden JSON blocks that never appear in visible text.

What Expand ships

According to the announcement:

• Iris, a vision model that reads the rendered page "like you do" and writes the main content as markdown that preserves icons, tables and layouts. Navigation, footers and banners go to an appendix that is one call away. Expand says it processes about 500,000 tokens per second and took a year to build; those are the company's statements.
• X-Ray, a second model that extracts useful data from a page's hidden JSON and filters the noise.
• Wayback Machine, a browsing history so you can see what your agent saw.
• Highlights, where you pass a query and get back only the passages that answer it, with cited snippets.
• Crawl across whole sites and WebSync for change tracking.

Both models run on every fetch, on the page as a real browser loads it, with JavaScript rendering, residential proxies and bot protection handled by the service. The announcement says engineers at OpenAI and Amazon Science use it, and the site lists Mixedbread among trusting engineers; we could not verify those customer claims.

The benchmark, carefully read

Setup from the page: 1,094 questions on 326 hard pages; Claude Sonnet 5.5 answers from each API's markdown at its default settings, graded strictly; tested October 8, 2026; "Expand is the only provider tested that includes the hidden JSON from the page, via X-Ray."

What the page says about MDBench Hard v1: it is "a set of hard web extraction cases that we curated together with our customers", spanning enrichment, procurement, pricing and competitive intelligence, and "it intentionally targets what existing web-to-Markdown systems get wrong". It is a question-answering benchmark where an agent gets only the markdown each API returns.

Our reading:

1. The intervals do not overlap, so within this benchmark the gaps are unlikely to be noise.
2. The benchmark is chosen to expose competitors' weaknesses. The page says so. That is a reasonable way to build a hard test, but it is not a neutral sample of the web.
3. Only Expand supplies hidden JSON, so the comparison is partly about a feature others do not offer, not only about markdown quality.
4. Competitor settings matter. "Default settings" for rivals may not be their best configuration.
5. No independent run. We found no third-party replication, and the benchmark does not appear to be public beyond the description.

Claimed versus verified

Pricing details

The pricing page says credits cost $1 per 1,000, one credit per successful fetch, with failed, blocked and empty requests costing nothing, and 10,000 free credits ($10) for new accounts. Highlights adds one credit per request. PDFs convert at one credit a page, scanned pages and images add five credits ($0.005) per page for the vision model, YouTube transcripts add five credits a video, and audio transcription is ten credits ($0.010) a minute. The free tier lists concurrency of 2. Compare that cost with the per-answer cost thinking in Microsoft's Token Wars microhack: a better fetch is only a win if the accuracy gain outweighs the extra spend on your tasks.

What people are asking

Why not just screenshot the page and give it to a vision model? You can, and for one-off pages it works. At scale it is slow and costly, and you lose machine-readable structure. Expand's approach is to run a purpose-built vision model on the render and output markdown, so the downstream model reads text but keeps the icon and table meaning.

What is "hidden JSON" and why does it matter? Many modern pages ship data inside script blocks or embedded state, for example product variants, prices by region or availability, that is not visible in the text a user reads. Extracting it can answer questions the visible markdown cannot. It can also surface data a site did not intend to show prominently, so consider the terms of use of what you fetch.

Does a better fetch tool remove the need for prompt-injection defences? No. Cleaner markdown does not make hostile page text safe. Any text your agent reads from the web can contain instructions aimed at it, which is why input from fetch tools should be treated as untrusted data.

Is Firecrawl at 27.2 percent really that bad? Not necessarily. The benchmark was built to target cases where ordinary conversion fails, so low scores for conventional converters are expected by design. On ordinary article pages the gap would likely be smaller, but Expand does not publish that comparison on the page we read.

Will this replace search APIs? They solve different problems. A search API finds pages; a fetch API reads one accurately. Expand lists Highlights and Crawl, which move toward retrieval, but Exa appears in the benchmark as a markdown provider, not as a search comparison.

How to test it yourself

1. Pick 30 to 50 pages your agent actually reads: pricing pages, comparison tables, listings, docs with icons.
2. Write questions with known answers, including ones that hinge on icons or table cells.
3. Fetch each page with your current tool and with Expand, using the free credits.
4. Give the same model the same markdown from each and grade strictly.
5. Record cost, latency and failures, not just accuracy.
6. Check what happens on pages that block bots, and read the terms for the sites you fetch; stealth and residential proxies raise legal and ethical questions of their own.

Why it matters

Agent reliability is often capped by input quality. A model with excellent reasoning, handed lossy markdown, will be confidently wrong. Tools that preserve structure, provide citations back to the page and keep a snapshot you can audit address a real gap, and the "every line links back to the exact spot on the page" feature is valuable in any workflow where someone must check the agent's source. It fits our wider view that agent skills and tools matter as much as the model, and that agents given the open web need guardrails.

Note also a naming echo: Expand describes Iris as "Jev for the web", referring to the Jev behaviour-classifier model we covered in Respan's Span-01 and Jev and that is used in agent routing pipelines.

Related reading

• Bot traffic flooding small sites: robots.txt and rate limits
• Indirect prompt injection explained
• Respan Span-01 and Jev
• What are agent skills?
• What is MCP?
• Microsoft Token Wars
• Official: Expand announcement and pricing

Details reflect Expand's published pages as of October 14, 2026 and may change.