Parallel introduced a Fast mode for its Search API on 21 August at $1 per 1,000 requests, which it positions as ten times cheaper than the default search in frontier models and five times cheaper than other search APIs, with about 700ms average latency. It claims first place on the Artificial Analysis Search Index for speed per task and third for intelligence, where its own Search Advanced product is the state of the art.
The reasoning is an explicit pairing argument rather than a quality claim. As intelligent models get cheaper, search has to follow, because a retrieval call that costs more than the model reasoning over its results inverts the economics of an agent loop. Parallel names GPT 5.6 Luna, DeepSeek V4 Pro and Qwen3.8 27B as the model class Fast mode is meant to sit beside, and points it at customer-support assistants, factual question answering and general-purpose agent workflows rather than deep research.
Splitting a product line by cost per call is the operational consequence of treating retrieval as its own tier, which is the separation argued in agents searching, fetching and browsing as distinct operations and the same pressure visible when GPT-5.6 prices came down. The agents hub tracks the cost layer.
The price and latency figures are Parallel's own, and the index positions are cited without the evaluation configuration. The condition to watch is whether the speed ranking holds at the cheaper tier under third-party testing, because a Fast mode that trades recall for latency moves the failure from the bill to the answer.