Expand Search is here.
Two weeks ago, we launched Expand Fetch, a page-reading API powered by models we trained from scratch. It preserves tables, icons and hidden data.
It's already handling millions of requests in production. But our users keep telling us they're frustrated with the search APIs they use alongside Fetch.
If your agent gets one search to answer a question, finding the right source is critical. When it's doing deep research, every query adds up.
Today we're introducing Expand Search, a web search API for agents with leading benchmark performance at $2 per 1,000 searches.
An animated example of Expand Search. The question "how do I rotate a Stripe API key" is entered automatically. Search returns three ranked pages from Stripe documentation, HowToRotate and Stripe Support, each with a title, excerpt and relevance score. Results are from a saved search.
How it compares
We compared Expand with seven other search providers on five public benchmarks covering coding questions, factual lookups and multi-step research. Our composite index averages answer accuracy across these public benchmarks, giving each equal weight.
Expand scored highest and was Pareto-optimal on our index. It delivered accuracy comparable to Exa Auto and Brave LLM Context at less than half the cost. No other configuration we tested matched its score for less money.
Expand Search / Benchmarks
Compare the configurations we tested across five datasets.
12 configurations · Five datasets, equally weighted · 2,985 questions
Index points, not answer accuracy. List price per 1,000 searches. Model costs excluded.
Frontier is an illustrative fit to these points.
Similar overall scores can hide very different tradeoffs. Some APIs reach comparable accuracy at several times the search spend. Others cost less per query but fall behind when the agent has only one search to work with. The comparisons below show where Expand's price buys you more.
Expand vs Exa and Brave
The price per 1,000 searches doesn't tell you what a research task will cost. Agents can search repeatedly as they investigate a question. Those calls add up.
On the 600-question AA-Omniscience benchmark, Expand reached 90.7% accuracy, with search spend of $6.16 per 1,000 questions and an average of 3.1 searches per question. Exa's and Brave's search APIs scored between 88.3% and 89.8%, with search costs ranging from $13.56 to $23.75 and averages of 3.2–3.9 searches per question.
Research accuracy for up to 74% less.
AA-Omniscience · 600 public questions
- $6.16ExpandAnswer accuracy 90.7%
- $13.56Exa LiteAnswer accuracy 89.5%
- $19.68BraveAnswer accuracy 89.8%
- $22.68Exa AutoAnswer accuracy 89.7%
- $23.75Exa FastAnswer accuracy 88.3%
Expand vs Perplexity Fast
Sometimes your agent has to answer after a single search. There's no follow-up query to fill in what it missed. In that setting, accuracy matters more than saving a fraction of a cent on the call.
Expand costs $0.002 per search, compared with $0.001 for Perplexity Fast. Expand reached 52.8% accuracy on SEAL-Hard's 254 questions, versus Perplexity's 45.3%. On the 824-question FRAMES benchmark, it scored 75.8% versus 72.8%. Across both benchmarks, the model answered 44 more questions correctly using Expand, a 6% increase for $1.08 in additional search spend. On WebCode, SimpleQA Verified and AA-Omniscience the two were level.
6% more correct answers. $0.001 more per search.
Across 1,078 questions · Compared with Perplexity Fast
SEAL-Hard · 254 questions
- Expand52.8%95% interval: 46.5 to 58.7 percent.
- Perplexity45.3%95% interval: 39.4 to 51.6 percent.
FRAMES · 824 questions
- Expand75.8%95% interval: 72.9 to 78.8 percent.
- Perplexity72.8%95% interval: 69.7 to 75.8 percent.
Expand vs Firecrawl
You don't need a bigger search budget to get more answers right. At comparable search prices, Expand outperformed Firecrawl on every benchmark we tested, by as much as 36 percentage points.
The biggest gap was on WebCode, where Expand reached 87.6% accuracy versus Firecrawl's 51.8%. Similarly, it scored 52.8% versus 41.3% on SEAL-Hard, and an F1 score of 97.6% versus 91.4% on SimpleQA Verified.
If you're using Firecrawl for search or page reading, switch to Expand Search and Fetch.
We are not the same
WebCode · 307 questions
Expand
269 / 307 correct
95% accuracy interval: 84 to 91.2 percent.Firecrawl
159 / 307 correct
95% accuracy interval: 46.3 to 57.3 percent.How we tested
We tested Expand Search against other providers in two settings: single-shot, where the model gets one search call before answering, and research agent, where it can search and fetch pages for up to 25 turns before submitting its answer.
Within each benchmark, all providers were evaluated on the same questions, using the same models and prompts to generate and grade the answers.
Single-shot
WebCode · SimpleQA Verified
FRAMES · SEAL-Hard
Search once, then answer.
Research agent
AA-Omniscience
600 public questions
Search and read. Up to 25 turns.
Benchmarks
| Benchmark | Questions | What it tests |
|---|---|---|
| WebCode | 307 | Finding documentation to answer coding questions. |
| SimpleQA Verified | 1,000 | Answering short factual questions. We evaluated it with search. |
| FRAMES | 824 | Combining facts from multiple Wikipedia articles. |
| SEAL-Hard | 254 | Reasoning with search results. |
| AA-Omniscience | 600 | Factual questions across six domains, answered by a research agent with up to 25 turns. |
Our benchmark writeups explain each protocol, the graders and how our setup differs from the original evaluations.
Scoring
For each provider, we calculate the percentage of questions answered correctly on each benchmark, then average those five percentages to produce the composite index. Each benchmark gets equal weight, regardless of how many questions it contains.
The WebCode tab displays evidence coverage, and the SimpleQA tab displays F1. These are separate metrics. The composite index uses answer accuracy for both.
Costs
The overall chart shows the price per 1,000 searches. Individual benchmark tabs show search spend per 1,000 questions, including every search the agent makes on AA-Omniscience. Model and grading costs are excluded. Prices are for the specific configurations we tested.
Evaluation details
Each configuration attempted the full question set for every benchmark, with one run between October 26 and 28, 2026. Questions left unanswered because of API errors after retries received no credit and remained in the total question count. We calculated 95% confidence intervals by repeatedly resampling the recorded per-question results. These intervals reflect variation across questions. They do not measure how results might change across repeated runs or different grading models.
These results come from our evaluation setup and are not directly comparable with the benchmarks' original leaderboards. Expand had the highest composite score. Its lead over Exa Auto, Exa Lite, Exa Fast and Brave LLM Context is within the margin of error. It is significantly ahead of the other seven configurations.
Evaluation code, prompts and per-question results are available in the benchmark repository.
Make the switch
Five searches cost one cent. Let your coding agent set up Expand Search, or get an API key and make your first search.
Search is our second product launch, following Expand Fetch. We're building toward a web your agents can query like a database. Join the waitlist for what's next.
Query the web like a database
Structured data from any site, as rows you can query. We'll email you when it opens.
About Expand
Expand is an applied research lab building web data APIs for AI agents. We design and train our own models from scratch.
