Artificial Analysis 用 Stirrup 測試 Parallel Search 的 fast 和 advanced 兩種模式,讓 GPT-5.6 Luna 用它搜尋網頁,完成 1,700 項搜尋任務。
- advanced 在 Artificial Analysis Search Index 得 83 分,與 Keenable Search(pro)並列第一。
- fast 得 82 分,排名第 4。兩種模式在 DeepSearchQA 都拿下 89 分 F1,名列前兩名。fast 平均每項任務約 $0.014,advanced 約 $0.030。
- 這表示搜尋工具也能靠近高分同類服務,同時維持較低費用。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
Parallel Search scores 83 on the Artificial Analysis Search Index, level with Keenable Search (pro) at the top of the board, with a fast mode at 82 for $0.014 per task
We benchmarked two modes of @p0's Parallel Search API, fast and advanced, in our Stirrup agent harness with Parallel Search as its web_search tool to complete 1700 search-focused tasks.
In our August runs, advanced scored 75 and fast scored 73. In these November runs, advanced scores 83 and fast scores 82.
Key benchmarking results for Parallel Search:
➤ Artificial Analysis Search Index: Parallel Search (advanced, objective) scores 83, 1st of 26, and Parallel Search (fast, objective) 82, 4th. That is a ~50-point lift over the same model with no search (33). Advanced is level with Keenable Search (pro) at 83 and 3 points ahead of Exa (auto, dynamic highlights) and Perplexity (medium) at 80. Fast is slightly ahead of both
➤ DeepSearchQA is Parallel's strongest benchmark: both modes score 89 F1, 1st and 2nd of 26, comparable to Keenable Search (pro) at 88. Advanced also ranks 1st on AA-Omniscience (74%), just ahead of Keenable Search (pro) (73%), and 3rd on BrowseComp (87%, level with Keenable Search (pro) and Perplexity (medium))
➤ Low cost: fast costs about $0.014 per task, 2nd lowest of 26 behind TinyFish ($0.009), against a board median of $0.063. Advanced costs about $0.030 per task, 6th lowest
Key details:
➤ Setup: GPT-5.6 Luna (medium) in Stirrup, Parallel Search API with mode=fast and mode=advanced
➤ Tool shape: the model fills Parallel's objective (optional) and search_queries (1 to 3) fields on each call
➤ Pricing: $1 per 1k requests (fast) and $5 per 1k requests (advanced)