Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
Removing Retrieval Fees Does Not Remove Reading Costs
TinyFish provides web search and fetching infrastructure for AI agents. In May 2026 it announced that Search and Fetch would be free: there is no per-request charge for Search or per-URL charge for Fetch, though both remain subject to usage limits. The offer addresses the service charge for retrieving live webpages, not the full computational cost of completing an agent task.
Search and fetch bring material back; the model still has to read it, reason over it, and produce an answer. Even when retrieval costs nothing, webpage text sent to the model consumes input tokens, and subsequent reasoning and output may also be billed. Treating the listed price of a search call as the total cost of a task misses how much the agent actually reads.
The Extractor Determines How Much the Model Has to Read
Fetch accepts a URL, renders the page, and extracts content as Markdown, JSON, or HTML. Removing navigation, ads, and cookie notices can keep irrelevant text out of the context. This moves some cost control upstream into retrieval: the less irrelevant material an extractor returns, the less the model may have to process. But shorter is not automatically better if it drops content the task needs.
On June 2, 2026, TinyFish evaluated five providers across 45 URLs using their default settings. For one Daily Mail article with a body of roughly 4,300 characters, TinyFish Fetch returned 4,737 characters, while Firecrawl returned 171,095—about 36 times as many. Exa Contents returned 65,793 characters and Tavily Extract 46,433. Parallel Extract returned just 600 characters, but that was an excerpt, not the full article. The comparison shows how much output size can vary; it is not a universal ratio across websites, and an excerpt is not equivalent to a full-page extraction.
Using the source's rough estimate of four characters per token and an illustrative input price of $1 per million tokens, TinyFish's output amounts to about 1,184 tokens, or $0.00118 in input processing; Firecrawl's amounts to about 42,774 tokens, or $0.04277. These estimates cover only input processing for this page, not retrieval, reasoning, or output, and actual token counts vary by tokenizer. The character counts come from TinyFish's own evaluation: useful for illustrating output size as a cost variable, but not enough to establish that one provider is better across sites or tasks.
Long Pages Consume Context and Can Reappear Across Turns
The input cost of one page may look small, but long contexts and multi-step agent flows change the calculation. In the source's estimate, a single Firecrawl extraction contains about 42,774 tokens—roughly one-third of a 128K context window. Three such pages total about 128,322 tokens, exceeding that window before system instructions, conversation history, or the answer are included. This assumes all extracted text enters the same context without filtering, truncation, or summarization, so it is a stress example, not a prediction of every agent task.
Agent loops may also resend earlier conversation turns. Artificial Analysis notes that each turn resends the full conversation, with earlier inputs billed at the model's cached-input rate. Caching can change the price of repeated input, but it does not make webpage content costless. The number of searches and fetches, and whether content is filtered before reaching the model, all affect total usage. Optimizing the price per million tokens alone cannot remove the expense of unnecessary context.
A Leaderboard Measures a Configuration, Not One Capability
An Artificial Analysis leaderboard snapshot dated September 28, 2026, displayed 12 configurations. TinyFish Search (web) scored 71, cost $9.30 per 1,000 benchmark tasks, and took 59 seconds per task, including model and search time. It had the lowest total cost among the configurations shown, but was the second slowest. Perplexity Search (medium) scored highest at 80; the no-search baseline scored 33. These figures describe the cost and performance of particular task, model, and retrieval combinations—not the search API in isolation.
TinyFish Agent uses a web-native model designed for search, retrieval, and interaction with live websites. TinyFish manages the model infrastructure and uses a cloud browser fleet to handle pages; the source says no local GPU is needed, but the Agent's built-in model cannot be swapped. For a team, that is a trade: managed integration can simplify deployment, while limiting model choice. Teams that must use their own model can assess free Search and Fetch as a separate retrieval layer. Teams considering the bundled Agent should compare model lock-in and task latency alongside cost.
Budget for Tasks, Not for the Word “Free”
The current pricing page lists standard limits of 30 Search requests per minute and 500 per hour, plus 150 Fetch URLs per minute and 1,000 per day. Within those limits, Search and Fetch do not draw from the Wallet. That boundary can be attractive for prototypes and moderate-volume workflows, but high-volume teams still need to check the limits and the rules that apply when they are exceeded. The available material does not give a complete cost path beyond the limits, so it does not support assuming retrieval stays free at any scale.
A practical evaluation should track request counts, returned characters or tokens, context length, repeated turns, model charges, and completion time on the team's own tasks. Then compare sending pages directly to the model with extracting or filtering the body first, while checking whether answer quality changes—not just the bill. TinyFish's page example shows the potential leverage of the extraction layer, while the leaderboard shows that low cost may come with longer waits. The right choice depends on whether the task needs the full page, how much latency it can tolerate, and whether the model must remain replaceable.