Custom AI Agents: What Web Reading Really Costs
TinyFish said on October 18, 2026 that its Search and Fetch tools are free within rate limits, while arguing that the bigger cost for custom AI agents often comes after retrieval, when page contents hit the model context. That matters for enterprise teams because a low search bill can still hide high token spend, slower loops, and context-window pressure in production. According to MarkTechPost’s coverage of the TinyFish release, the company’s core point is that teams often budget for the model but underestimate the cost of retrieved content.
TinyFish makes search free, but the agent bill is still real
The headline is straightforward: TinyFish is making Search free at up to 30 requests per minute and 500 per hour, and Fetch free at up to 150 URLs per minute and 1,000 per day. On paper, that changes the retrieval line item for teams building AI automation agents that read the web.
But the operating cost does not disappear. It moves downstream. If a fetched page adds thousands or tens of thousands of tokens to the context, the model still bills for input processing, reasoning, and output. For enterprise AI integrations, that distinction matters more than the headline price.
MarkTechPost paraphrases TinyFish’s position as treating search like infrastructure inside execution, similar to storage or DNS for ordinary software. That is a useful framing, but it also highlights the design challenge: if search becomes cheap, payload control becomes the new budgeting problem.
Retrieval charges are only the first line item
For most AI agent development projects, web reading creates two costs. The first is retrieval: search requests, fetch calls, or extract calls. The second is model processing: every returned page that enters the prompt consumes tokens, and every reasoning step or final answer adds more.
That is why model routing is only a partial solution. The article points to Stripe’s agreement to acquire OpenRouter, a signal that enterprises increasingly route workloads across many models to lower token prices. Routing can reduce price per token. It does not reduce the number of tokens created when a scraper returns a bloated page.
This is the implementation issue many teams miss when planning AI integration services. Procurement may compare vendors on request pricing, while engineering absorbs the hidden cost later in prompt size, retries, summarization steps, and slower task completion. Teams evaluating AI automation implementation services usually need to budget both layers together, not as separate buying decisions.
One article can range from 1,184 to 42,774 tokens
The strongest operator lesson in the source material is not that one vendor is cheaper than another. It is that extraction shape changes total cost dramatically.
In TinyFish’s June 2, 2026 fetch evaluation, one Daily Mail page of roughly 4,300 characters produced very different outputs depending on provider. TinyFish returned 4,737 characters, or about 1,184 estimated tokens. Exa returned about 16,448 estimated tokens. Firecrawl returned about 42,774 estimated tokens. Parallel returned only an excerpt, around 150 estimated tokens, which makes it cheaper to process but not directly comparable to full-page extraction.
A simple way to read those numbers: if an AI API integration passes every extracted character into the same prompt, one vendor’s output may be manageable while another’s consumes a large share of the context window before the agent has even started reasoning. That is not a theoretical issue. It changes run cost, latency, and answer quality.
The source also notes an important limit: this comparison comes from a single webpage and should not be generalized across all sites. Still, the directional lesson is hard to ignore. Payload discipline can matter more than nominal search price.
For third-party benchmark context, the Artificial Analysis Search API leaderboard held the answer model fixed and varied the search provider. On the September 28, 2026 snapshot cited by MarkTechPost, TinyFish posted the lowest displayed total cost at $9.30 per 1,000 benchmark tasks, but also the lowest displayed Search Index score at 71.
Why large page returns squeeze the context window
A 128K context window sounds generous until an agent starts reading raw web pages. Using the source article’s math, a 42,774-token extraction would consume about one-third of that context. Three similarly large pages would total roughly 128,322 tokens, leaving little or no room for system instructions, prior turns, or the model’s answer.
This is where custom AI agents tend to get unexpectedly expensive. ReAct-style loops, browser retries, and multi-step research tasks often resend conversation state. Artificial Analysis explains its methodology by noting that each turn resends prior context, with earlier input billed at cached-input prices when available. Cached input lowers reread cost, but it does not eliminate it, and it does not solve context crowding.
The operational implication is non-obvious but important: large retrieval payloads create a double penalty. They raise token usage and reduce the room available for control instructions, making agents more likely to truncate, summarize poorly, or make another call to recover missing detail. In SaaS, fintech, and e-commerce workflows, that can turn into slower customer support actions, noisier research agents, or more brittle back-office automations.
What the leaderboard says about speed, score, and price
The TinyFish story is credible precisely because it includes trade-offs. Free retrieval within limits is real. Lower displayed total benchmark cost is real. But so are the weaker benchmark score and slower task time in the cited snapshot.
According to the MarkTechPost summary of the September 28, 2026 leaderboard, TinyFish Search (web) scored 71 at $9.30 per 1,000 benchmark tasks and 59.0 seconds per task. Perplexity Search (medium) scored 80, while the no-search baseline was 33. Exa, Firecrawl, and Parallel all posted higher benchmark scores than TinyFish in the displayed set, though at materially higher total cost.
That creates a practical buying framework for AI implementation services:
- If the workflow is cost-sensitive and can tolerate lower search quality, free retrieval plus tight prompt controls may be attractive.
- If the workflow depends on higher recall or better benchmark performance, a higher-cost provider may still be the better fit.
- If the workflow needs a swappable model inside the vendor’s agent product, TinyFish’s built-in model constraint becomes a real consideration.
It is also worth separating benchmark cost from API list price. The TinyFish pricing page describes request and usage pricing, while Artificial Analysis measures cost per 1,000 benchmark tasks. Those are not the same unit because one task may involve multiple searches, repeated reasoning, and variable payload sizes.
When free retrieval is actually the right choice
Free retrieval is most useful when teams already know how to constrain extraction, summarize early, and discard low-signal page content before it reaches the expensive part of the stack. In other words, search price matters most when the implementation around it is disciplined.
That makes this less of a vendor-pricing story and more of an architecture story. The teams that benefit most are the ones treating web access as one component inside a broader agent system: payload filters, max-token caps, context management, fallback logic, and model routing all have to work together.
What to watch next is whether benchmark leaders close the cost gap without inflating payload sizes, and whether buyers start asking for token-per-task reporting instead of just request pricing. As more enterprise teams move from experiments to production custom AI agents, that reporting layer may become the more important competitive feature.
Martin Kuvandzhiev
Co-Founder & CEO, encorp.ai
CEO and Founder of Encorp.io with expertise in AI and business transformation
LinkedIn