TinyFish Makes Search and Fetch Free for AI Agents, Pointing to Token Volume as the Real Cost

According to a MarkTechPost report on Oct. 18, 2026, the company's case rests on a two-part bill: the retrieval charge that a search or fetch API bills per request, and the model processing cost incurred when the returned content is passed into an AI model as input tokens, with reasoning and generated output billed on top. TinyFish says teams budget for the model but underestimate how much retrieved content drives that spending.

The free tier covers 30 search requests per minute and 500 per hour, plus 150 fetched URLs per minute and 1,000 per day. Inference on the returned text is still paid to whichever model processes it, so the size of each returned page matters as much as the number of calls.

On the Sept. 28, 2026 snapshot of the Artificial Analysis Search API leaderboard, TinyFish Search (web) scored 71 at $9.30 per 1,000 benchmark tasks, with 59.0 seconds per task including model and search time. The company said that was the lowest total cost among the 12 configurations displayed. Perplexity Search (medium) posted the highest displayed score at 80, while the no-search baseline scored 33. TinyFish also recorded the second-longest time per task among the displayed configurations, and its Agent product uses a built-in model that cannot be swapped.

TinyFish ran a fetch quality evaluation on June 2, 2026, across 45 URLs and five providers on default settings. For one Daily Mail article with a body of roughly 4,300 characters, TinyFish Fetch returned 4,737 characters, or about 1,184 tokens, against Exa Contents at 65,793 characters (about 16,448 tokens), Firecrawl Scrape at 171,095 characters (about 42,774 tokens) and Tavily Extract at 46,433 characters (about 11,608 tokens). Parallel Extract returned a 600-character excerpt rather than the article. The token and cost figures are MarkTechPost calculations at roughly four characters per token and an illustrative input price of $1 per 1 million tokens, covering input processing only.

In that example Firecrawl returned about 36 times as many characters as TinyFish, and Parallel's excerpt is not directly comparable to a full-page extraction. The company cautioned that the comparison covers a single webpage and should not be generalized across all sites.

Page size feeds directly into context limits. At about 42,774 tokens, one such page would occupy roughly a third of a 128K-token context window; three would take about 128,322 tokens, leaving little or no room for system instructions, conversation history or the agent's response. Agent loops resend context as well. Artificial Analysis notes that each turn resends the full conversation, with earlier input billed at the model's cached-input price. Caching lowers the cost of rereading a noisy page but does not remove it, and the page still occupies the window.

In the same internal evaluation, a result counted as covered only if it returned at least 1,000 characters and scored a judged signal ratio above 30%. Coverage was 93% for TinyFish, 80% for Tavily, 73% for Exa, 62% for Firecrawl and 58% for Parallel. Among usable results only, median signal was 90.5% for TinyFish and 89.7% for Exa. Outputs above 30,000 characters were truncated before judging. TinyFish ran the test on itself, the sample was 45 URLs, and news pages drove most of the gap.

The pricing move comes as enterprises route each task to the cheapest capable model, a trend underlined when Stripe agreed to acquire OpenRouter, a gateway spanning more than 400 models from more than 80 providers. Routing lowers the price per token but does not change how many tokens a retrieved page contains, which is why TinyFish argues that page size, not the per-request fee, is the number teams should watch. The company frames search as one operation inside execution, like DNS or storage for ordinary software: an agent booking, comparing or filling out a form should care whether the job got done, not which search system found the page.