According to CoinMeta, OpenRouter has launched web, search, and benchmarks to test the networking and search configuration of AI. The first batch covers search engines such as exa, parallel, and perplexity, and is paired with models like gpt-5.6, SOL, claude, opus, deepseek, v4, and flash to compare accuracy, cost, and speed. The results show that the number of search rounds is more important than switching search engines. In browsecomp difficult questions, increasing the search limit from 1 round to 25 rounds doubled the score, but increased the cost per question by 2.5 to 7 times. For example, when claude and opus were paired with perplexity, the accuracy increased from 35.8% with 1 round to 89.0% with 25 rounds. The model itself is also more important than the search engine; switching models can average an increase of about 15 points, while switching to search engines like exa, parallel, and perplexity averages an impact of about 10 points. It is worth noting that when the answer is correct, the model searches on average 10.3 times, and when it is wrong, it searches 19.7 times. Allowing more search attempts may lead to higher failure costs.