Editorial illustration for TinyFish Free Search: 90.5% Signal Tops Exa 89.7% as News Drives Gap
TinyFish Search Beats Exa on Cost-Efficiency
TinyFish Free Search: 90.5% Signal Tops Exa 89.7% as News Drives Gap
TinyFish put a number on something most AI teams never itemize: the cost of letting an agent read a live webpage. On the Artificial Analysis Search API leaderboard's September 28, 2026 snapshot, TinyFish Search (web) scored 71 at $9.30 per 1,000 benchmark tasks, taking 59.0 seconds per task for model plus search time. That score sits at the bottom of the 12 configurations shown, below Perplexity Search (medium) at 80.
But TinyFish isn't selling on accuracy alone. Within rate limits, 30 requests a minute and 500 an hour for search, 150 URLs a minute and 1,000 a day for fetch, retrieval itself is free. No GPU, no local model to host.
The company runs its own web-native model on managed infrastructure, with a cloud browser fleet doing the actual page rendering, reachable through REST, Python, TypeScript, a CLI, or MCP. The pitch is that teams price out the model and forget the bill that comes from everything the model has to chew through first: the tokens spent on retrieved content, not just the tokens spent reasoning about it. What that retrieval cost actually looks like, and why it doesn't show up on a leaderboard, is where the math gets interesting.
Free Search and Fetch remove TinyFish’s retrieval charge within its limits. The downstream model still incurs inference costs. That is why the size of each returned page matters.
Why this matters A 45-URL sample and a 0.8-point signal gap isn't evidence of a durable quality edge, it's a company grading its own homework. TinyFish ran this test on itself, against Exa, and the margin narrows to statistical noise once you account for sample size and the fact that news pages, not the broader web, drove most of the difference. That matters for anyone evaluating retrieval vendors: a benchmark built by the vendor being benchmarked deserves the same scrutiny you'd give a chipmaker's own power-efficiency claims.
The more interesting idea here isn't the 90.5% number, it's the framing of search as execution infrastructure, priced like DNS or storage rather than billed per call. If that pricing model holds up at scale, it shifts where AI agent costs actually land, from retrieval fees to token processing, which founders budgeting agent infrastructure should factor in now rather than after they've built around per-request pricing. Worth watching: whether "free" retrieval survives contact with real traffic volumes, and whether independent benchmarks replicate that signal gap outside TinyFish's own test set.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Anthropic Nearly Stops Junior Hiring as Claude Agents Log Six Times Hours
From Archives Anthropic consults law firms to set fair negotiation norms for AI agents
Common Questions Answered
How does TinyFish's pricing model differ from competitors like Perplexity Search on the Artificial Analysis Search API leaderboard?
TinyFish Search scored 71 at $9.30 per 1,000 benchmark tasks, positioning it below Perplexity Search (medium) which scored 80. TinyFish's competitive advantage lies not in raw accuracy scores but in its free retrieval capabilities within rate limits, offering 30 requests without retrieval charges, though downstream model inference costs still apply.
What does TinyFish's Free Search and Fetch feature actually eliminate from AI agent costs?
TinyFish's Free Search and Fetch remove the company's retrieval charge within specified limits, but the downstream model still incurs inference costs. The size of each returned page directly impacts these remaining costs, making page retrieval efficiency a key factor in the overall expense calculation for AI agents reading live webpages.
Why should the 0.8-point signal gap between TinyFish and Exa be viewed with skepticism according to this analysis?
The benchmark was created by TinyFish itself and tested against Exa, raising concerns about objectivity and self-grading bias. Additionally, the 45-URL sample size is small, and the performance difference was primarily driven by news pages rather than the broader web, making the results statistically unreliable for general vendor evaluation.
What is the actual performance cost in terms of speed when using TinyFish Search according to the September 28, 2026 leaderboard snapshot?
TinyFish Search took 59.0 seconds per task when combining model inference time plus search time on the Artificial Analysis benchmark. This timing metric is important for AI teams evaluating the practical overhead of allowing agents to read live webpages in production environments.
Further Reading
- Best Web Search APIs & MCPs for AI Agents 2026 - Vellum
- Agentic Search Explained: How AI Agents Use the Web in 2026 - TinyFish
- 7 Best AI Search Engines for Developers & AI Agents (2026) - TinyFish
- TinyFish launches the Data Partners Alliance, opening the licensed web to AI agents - Yahoo Finance
- 8 Best Exa Alternatives & AI Competitors in 2026 - TinyFish
