You.com Web Search Highlights Reaches 95.17% on SimpleQA

You.com on September 1, 2026 introduced a new extraction mode for its Web Search API called Highlights, reporting a 95.17% score on the SimpleQA benchmark and a 35% lower total query cost than Claude Sonnet 5’s built-in web search in an independent evaluation. The feature is live through the API’s extraction_mode parameter at the same $5 CPM pricing as the existing service, the company said.

What Highlights Does

The Web Search API returns structured web and news results for AI applications. By default, each result carries a snippets array of short, keyword-centered text fragments. Setting extraction_mode to highlights replaces those snippets with a contents.highlights array containing the passages from each page that address the specific query.

According to You.com, the extraction runs per request, with each query triggering a fresh pass against the page. A model identifies which spans of text already answer the query and returns them verbatim, preserving numbers, dates, and figures as they appeared in the source. Tables return with headers intact, code blocks keep their formatting, and sentences from the same section that both matter are merged into a single passage. Candidate passages are ranked, with the highest-ranked returned first.

Accuracy Results

You.com said it ran three benchmarks across all three extraction modes. Highlights led on SimpleQA at 95.17% and on RetrievalQA at 66.93%, the two benchmarks built around single-fact lookups. On FRAMES, a multi-hop reasoning benchmark, full-page extraction scored 82.77% against Highlights’ 82.04%, a 0.73-point gap the company described as inside the run-to-run variance it observes on that test.

The company noted the accuracy gain is not free: cost per question rises about 11% versus Snippets because Highlights sends more text to the answering model. Cost per correct answer, which divides cost by accuracy, still comes out about 5% lower, it said.

Against external systems on SimpleQA, You.com reported three systems clearing 95%: You.com with Highlights at 95.17% and 695ms p50 latency, Exa (full page) at 95.47% and 1337ms, and Parallel (advanced) at 95.33% and 2098ms. The company said it displayed each competitor’s best-performing configuration, so the comparison favors them on setup.

Independent Cost Evaluation

Braintrust, as part of an independent evaluation, ran You.com Web Search with Highlights against the search tools bundled into the OpenAI and Anthropic APIs on 1,329 questions, with cost including both inference and search.

In that evaluation, Claude Sonnet 5 cost $0.0747 per question with Highlights against $0.1148 for its built-in search, a 35% reduction, at slightly higher accuracy (79.23% versus 78.78%) and 1.26 seconds faster. GPT-5.6 Terra cost $0.0417 per question with Highlights against $0.0467 built-in, an 11% reduction, with 3.66 points higher accuracy. Cost per correct answer dropped 35% for Claude and 15% for GPT, since Highlights answered more questions correctly on the same or lower spend, according to You.com’s account of the results.

Observability Advantage

You.com argued that unbundled web search offers better observability than the built-in tools from OpenAI and Anthropic. Built-in search returns the queries it ran and the pages it cited but not the passages the model reads, the company said, leaving no way to see which step produced a wrong answer.

It cited a question from Braintrust’s evaluation asking how many service games Carlos Alcaraz lost across two tournament runs combined, which required adding two separate numbers. The You.com configuration returned one passage establishing the first value at 24 and another establishing the second at 22, and the model added them to answer 46. The runs that missed never surfaced a passage supporting 24, used 22 for both values, and answered 44. Every run performed the arithmetic correctly, a fact the company said is only knowable because the passages appear in the trace.

When Not to Use It

You.com said Highlights adds about 160ms over Snippets on single-shot lookups, and that under a hard latency budget on simple Q&A questions Snippets remain the right default. For pages that change faster than an index refreshes, extraction_mode: "full_page" fetches live rather than serving from cache. FRAMES is where that trade-off shows up, the company said, because its multi-hop questions require a wider slice of context than a handful of tightly scoped passages provides, and full page beats Highlights there by 0.73 points.

The company said new accounts receive $100 in credits and that the harness producing its benchmark numbers is publicly available.