I ran Firecrawl, Exa, Parallel and Claude’s native web search against OpenAI’s SimpleQA benchmark to see how much of a difference the search provider actually makes. All four were tested with the same setup: a GPT-5.4 agent using high reasoning effort, with a maximum of 20 search or extraction calls per question. The answers were then graded by GPT-5.4 using OpenAI’s official SimpleQA grading prompt. Each provider was tested twice and I kept the better result. For comparison, GPT-5.4 without...