TinyFish Makes Search Free: What It Really Costs an AI Agent to Read the Web
MarkTechPost Asif Razzaq
TinyFish made Search and Fetch free within limits. That cuts one bill, but the model still pays for every page it reads.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
TinyFish is trying to change how people think about web search for AI agents: not as a separate bill, but as part of the job itself. Its Search and Fetch products are free within published limits, and the company says that’s possible because it built the browser fleet, session handling and scaling first, then layered products on top.
That does not make web reading cheap in the way marketing copy likes to imply. The retrieval charge can disappear, but the text still has to flow into a model, where it becomes input tokens. Reasoning and generated output add more cost after that. So the real question is not just what search costs. It’s how much content the search system sends back, and how much of that content the model has to chew through.
MarkTechPost points to a TinyFish fetch evaluation from June 2, 2026, across 45 URLs and 5 providers. On one Daily Mail article of about 4,300 characters, TinyFish Fetch returned 4,737 characters, or roughly 1,184 tokens by the article’s estimate. Exa returned 65,793 characters, Firecrawl 171,095, and Tavily 46,433. Parallel returned only an excerpt, about 600 characters, so it is not directly comparable to a full-page extract. The point is simple enough: one page can turn into very different token bills depending on the provider.
That matters because context windows are not bottomless. The article says a page with about 42,774 tokens would eat around one-third of a 128K-token window, and three such pages could push past that limit before there is room left for instructions or the agent’s reply. Repeated agent loops make this worse, even when caching trims the price of rereading earlier content.
Artificial Analysis gives TinyFish a mixed result. On its Sep 28, 2026 snapshot, TinyFish Search (web) scores 71 and shows $9.30 per 1,000 benchmark tasks, with 59.0 seconds per task. That is the lowest total cost among the 12 configurations shown, but also the lowest displayed Search Index score. The lesson here is not that free search is fake. It’s that the free part is only one piece of the invoice.
My take — AI-written commentary, not fact-checked reporting
The industry loves calling search “free” right up until the token meter starts spinning. TinyFish is being refreshingly blunt about that, which is more honest than the usual cloud fairy tale. The broader pattern is clear: the real moat is not a cheaper search box, it’s who controls the retrieval shape before the model gets involved.
Read more about this at: MarkTechPost
