What Is the Best LLM for Web Scraping?

The thousands-separator trap

Every example-page price was under £1,000, so a price regex that ignores comma separators passes every test the model can run. Three held-out products cost more. On £1,249.00, GPT-5.6 Terra's regex ([0-9]+(?:[.,][0-9]{1,2})?) captured 124 and Claude Sonnet 5's ([0-9.]+) captured 1. Both returned well-typed floats that were silently wrong. Grok 4.5 and GPT-5.6 Luna captured the full number and stripped the comma.

The positional-selector trap

The release date lives in a spec list. On all 8 example pages it is the first <dd>, so Grok 4.5, GPT-5.6 Luna, and Claude Sonnet 5 all wrote the positional selector dl.spec dd. On 11 of the 32 held-out pages, extra spec rows push the date out of first position and the selector silently returns a material or capacity string instead. GPT-5.6 Terra was the only model to anchor on the label itself, using //dt[normalize-space()='Released']/following-sibling::dd[1], and it did so without ever seeing a page that required it. On this evidence, that is the single most valuable authoring decision in the benchmark, and it is why Terra posted the best robustness delta (+0.009).

• Model selection: four models from the tier below the current frontier. We left out GPT-5.6 Sol, Fable 5, and Opus 5 because a benchmark all of them pass tells you nothing about what to run in production.
• Task: author a two-stage workflow on Parseium via MCP: crawl paginated listing pages of a server-rendered demo shop, follow to product detail pages, extract 11 typed fields (price, SKU, stock, GTIN, release date, and more).
• Identical inputs: one example URL and a 1,079-byte prompt, byte-identical across all four runs, reasoning effort set to high everywhere.
• Held-out scoring: we scored each workflow against a separate deploy of the shop (4 listing pages, 32 detail pages) with zero links between the example site and the scoring site.
• Isolation, audited: we grepped every session transcript for all 32 held-out product slugs and the scoring site's URL prefix: zero hits for every model.

That audit exists because our first attempt failed it. In the original batch, the scoring pages were reachable by pagination and Grok 4.5 found them. It browsed a scoring page, discovered the extra spec rows, and wrote the label-anchored selector, scoring a perfect 100. With the leak closed, its score dropped to 99.14 and it wrote the same positional selector as everyone else. The other three scores were identical to two decimal places across both batches. An unreachable held-out set changed both a number and a conclusion, which is exactly why we score only on pages the models cannot see.

• One run per model: the 0.94-point spread between first and last is within plausible single-run variance. We are running repeats before treating the ordering as settled.
• One target, and an easy one: a server-rendered shop does not test the decision that separates scraping agents most: when to pay for browser rendering. That harder target is next.
• Quality is saturated: all four models sit above 98, and three are within 0.1 of each other. This benchmark is a competence floor, not a fine-grained ranking.
• Cross-harness numbers are not comparable: Claude Sonnet 5 ran under Claude Code, so its token cost and wall clock reflect that harness as much as the model.