The Web Search LLM Leaderboard ranks 10 AI models by how well they answer 4,703 hard research questions with web search, drawn from BrowseComp, DeepSearchQA, HLE and Firecrawl Search Bench.
Average search lift with Firecrawl Alexandria: +40.6 points.Average search lift with Firecrawl Alexandria, open-weights models: +37.4 points.Average search lift with Firecrawl Alexandria, closed models: +43.8 points.
| # | Model | With Firecrawl Alexandria | Lift | Without | Cost per 1K | Time per question | Failed |
|---|---|---|---|---|---|---|---|
| 01 | Claude Opus 5.5Anthropic | 79.0 | +41.9 | 37.1 | $372 | 126 s | 0.2% |
| 02= | GPT-6.1 SolOpenAI | 77.9 | +45.6 | 32.3 | $208 | 134 s | <0.1% |
| 02= | Kimi K3.1 ProMoonshot AI | 77.4 | +38.9 | 38.5 | $527 | 386 s | 2.6% |
| 02= | Pareto 26.9Unbiased | 77.1 | +44.7 | 32.5 | $658 | 328 s | 0.2% |
| 05= | DeepSeek V4.1 Flash (1017)DeepSeek | 68.9 | +38.0 | 30.9 | $289 | 368 s | 0.7% |
| 05= | Mercury 3*Inception | 68.7 | +42.4 | 26.4 | $242 | 214 s | 0.1% |
| 07 | MiniMax M3.1 FlashMiniMax | 64.8 | +40.5 | 24.3 | $205 | 407 s | 0.2% |
| 08= | Claude Haiku 5.5Anthropic | 61.9 | +44.4 | 17.5 | $165 | 156 s | 0.1% |
| 08= | GLM 5.3 FlashZ.ai | 61.9 | +36.7 | 25.2 | $173 | 755 s | 2.5% |
| 10 | Nemotron 3.5 LightningNVIDIA | 44.0 | +32.8 | 11.2 | $89 | 111 s | <0.1% |
Also tested: Claude Fable 5.2, on 3 of 4 question sets so far, so it is not ranked here. See the full report →
Without tools, a model answers from its training data. We ask each one the same hard questions with and without Firecrawl Alexandria and measure how much its score rises.
Each point is one model: its score with Firecrawl Alexandria against what it costs.
Cost per 1,000 questions, log scale (model tokens + Firecrawl credits)
Up and to the left is better. Orange points sit on the frontier, where no cheaper model scores higher. The ringed point is the best-value pick above.
Firecrawl's library of data for AI agents: 400+ data providers plus Firecrawl's own Research, Developer and Government indexes. With a Firecrawl API key on a team that has Firecrawl Alexandria enabled, adding "alexandria" to the search sources returns them as tools inside a search call. In these runs, models with Firecrawl Alexandria could search the web, read pages and call those data tools. Learn more →
OpenAI
Obscure facts with one short answer, found by chaining clues across many searches.
BrowseComp paper“Give me the title of the scientific paper published in the EMNLP conference between 2018-2023 where the first author did their undergrad at Dartmouth College and the fourth author did their undergrad at University of Pennsylvania.”
Multi-step research tasks in 17 fields; about two thirds of the answers are lists.
DeepSearchQA paper“Between 2019 and 2020, which states had a death rate (the number of deaths per 100,000 total population) between 229.8 - < 268.1 that was inflicted by the leading cause of death in the US during 2019 and 2020 per the CDC?”
Center for AI Safety and Scale AI
Expert exam questions in mathematics, science and the humanities. Text-only scores aren't comparable to published full-HLE scores.
HLE website“Hummingbirds within Apodiformes uniquely have a bilaterally paired oval bone, a sesamoid embedded in the caudolateral portion of the expanded, cruciate aponeurosis of insertion of m. depressor caudae. How many paired tendons are supported by this sesamoid bone? Answer with a number.”
Firecrawl, November 2026
Private questions, mostly about pages published just before the set was built.
“In its October 2026 release notes, an open-source web framework changed its default build tool. Which tool replaced the old default, and in which version did the change ship?”