Behind TinyFish's Free Search: The Real Cost of AI Agents Reading the Web

Search APIs are now free—but does that really make it zero-cost for AI Agents to read the web? TinyFish's approach is worth noting, but there's an easily overlooked part hidden in its cost structure.

Free Search Is Only Half the Cost

According to MarkTechPost, TinyFish is a web retrieval infrastructure provider that recently made its search capability free to use. As the report breaks it down, an AI Agent reading the web actually bears two types of expenses: first, retrieval fees—the per-request unit price charged by search or scraping APIs; second, model processing fees—the tokens consumed by the body content from retrieved pages entering the model, plus the additional overhead of reasoning and generating responses.

TinyFish drives the first type of cost to zero, effectively covering the "entry fee" for developers. But the second type isn't determined by the retrieval provider—it depends on how long the page content is, how large the context sent into the model is, and how many rounds of reasoning the Agent runs. In other words, free search lets Agents initiate retrieval more frequently, which may actually push more web content into the model and shift the bill from the API side to the model side.

The report's wording is fairly measured, presenting only this cost framework without disclosing TinyFish's specific pricing strategy, free-tier caps, or technical implementation details. That means "free search" is currently closer to a pricing move than a complete technical cost revolution.

Tokens Are the Hidden Bill for Agents Reading the Web

In a typical Agent workflow, retrieval is only the first step. What really eats the budget is what follows: the retrieved HTML needs cleaning, the body text gets spliced into the prompt, the model must repeatedly reference this content across multiple rounds of reasoning, and finally it generates an answer. The news report explicitly points out that content passed from retrieved pages into the model consumes tokens, and reasoning and generation add new costs on top.

This structure has direct engineering implications for developers. The same search returning ten candidate links versus three puts completely different pressure on the model's context; for the same page, grabbing the full HTML versus only the body paragraphs can mean a several-fold difference in token consumption. After search becomes free, constraints on retrieval frequency are relaxed, but context windows and model call counts still have hard limits—the Agent's cost curve will be determined more by content processing strategy than by the per-request price of search.

From a cost-control perspective, free search is more like shifting the optimization focus from "call the API less" to "feed fewer tokens." Developers need to focus on page trimming, content summarization, and cache reuse rather than simply increasing retrieval frequency.

Context Windows and Token Billing: The input length of a large language model (LLM) is limited by the "context window"—the total number of tokens the model can receive in a single request (including system prompts, conversation history, retrieved content, and user input). A token is the smallest unit of text the model processes, roughly equivalent to a word fragment in English or 1–2 characters in Chinese. Mainstream APIs price input tokens and output tokens separately, and costs rise significantly when input token volume is large. Therefore, before splicing web page body text into the prompt, an Agent typically performs HTML cleaning, body extraction, truncation, or summarization to compress the number of tokens sent to the model. The larger the context window, the more retrieved content it can hold—but the higher the unit price, which is the underlying reason for the "free retrieval, paid model" cost structure.

HTML Cleaning and Body Extraction: Raw web page HTML contains a large amount of navigation bars, ads, scripts, and styles, with the actual body often accounting for only a small portion. Sending HTML directly into the model not only wastes tokens but also introduces noise that interferes with reasoning. Common approaches include Readability-style algorithms, DOM parsing, or dedicated body extraction services to reduce the page to plain-text paragraphs. Cleaning quality directly determines token efficiency: too loose and the context balloons; too tight and key information may be lost. This is also where retrieval providers like TinyFish can build differentiation beyond "free search"—scraping quality and content cleaning capability, rather than the retrieval request itself.

Impact on Agent Infrastructure Competition

Search retrieval has long been a paid link in the Agent toolchain, and per-request billing keeps the marginal cost of high-frequency retrieval Agents persistently high. By making this link free, TinyFish will lower the barrier to entry for Agent products at the infrastructure level and put pricing pressure on similar retrieval services.

But the report doesn't say whether the free offering comes with usage conditions, nor does it provide any quantitative data on model-side costs. So the only judgment that can be confirmed right now is this: once retrieval fees are removed, token processing costs will further rise as a share of an Agent's total overhead, becoming a more visible cost item. Whether TinyFish can build a long-term advantage with this strategy depends on how much added value it can provide going forward in scraping quality, content cleaning, and context optimization—exactly the links that determine the level of the second type of cost.

For teams building Agents, what's actionable now isn't rushing to switch retrieval providers, but first getting a clear picture of their own token bill: how much of each task round's context actually comes from retrieved pages, and how much of that content is truly used by the model. Only once this is clear can the benefits of free search be realized.

Agent Toolchains and Per-Request Billing: An AI Agent is typically driven by a multi-round loop of "plan—tool call—observe—re-plan," with external capabilities like search, scraping, and code execution connected as tools. Retrieval tools have long used per-request or per-scrape billing, and the marginal cost of high-frequency retrieval Agents grows linearly with call volume. Making retrieval free is equivalent to moving the charging point from the "entrance" to "model processing," which will change Agent products' cost models and pricing strategies and force similar retrieval providers to reassess their own charging points and added value.

One-line Summary

TinyFish's free search only eliminates retrieval costs—the real bill for Agents reading the web lies on the token processing side.

Paragraph Summary

TinyFish makes web search free, seemingly eliminating the cost for AI Agents to read the web. But the report's breakdown shows that Agents reading the web actually bear two types of expenses: per-request retrieval API fees, and the tokens and reasoning/generation costs consumed by the model processing retrieved content. Free search only lowers the first type—the "entry fee." The second type is determined by page length, context size, and reasoning rounds, and is not controlled by the retrieval provider. More importantly, free retrieval may encourage Agents to search more frequently, pushing more web content into the model and shifting costs from the API side to the model side. The report doesn't disclose TinyFish's pricing details or free quota, so this is currently more a pricing move than a technical cost revolution. Developers should shift their optimization focus from reducing API calls to page trimming, content summarization, and cache reuse.