Introducing TinyResearch and TinySet: Delegate the deep web work

We didn't set out to build a research product or a dataset product. We noticed our own enterprise team building both, by hand, for customer after customer.

One week it was a proper research report on a market someone was about to enter. The next it was a dataset that didn't exist anywhere in one piece, pulled together from dozens of sites. Every time, it was the same work stitched together from our existing primitives, and every time it took longer than any of us would like to admit.

At some point it got a little silly to keep doing it as a one-off. So today we're launching both as products: TinyResearch and TinySet.

If you'd rather see them than read about them, we put many reports and datasets, with the exact prompts behind them, at tinyfish.ai/deep-work.

Why these two, and why together

If you've ever had to really understand a market, a company or a decision, you know the work usually comes down to two jobs. Sometimes you know exactly what you need and it's the same few facts across a lot of things, like what the vendors in a category charge, their security and privacy links and certifications, or which data centers are being built where. Other times it's one hard question that needs a proper investigation, with sources you can check and an answer laid out so you can actually decide on it. Or, honestly, something scannable with the right data to feel confident in your decision making process.

Both of those eat whole days, and honestly, the AI tools people hand them to today don't help much. Ask for a table and you get a dozen rows and still have to go find the rest yourself. Ask a hard question and you get a wall of text you have to dig through before any of it is usable.

So we built a product for each job. TinySet collects the same facts across many companies, places or products and gives you a dataset, and TinyResearch takes one hard question, investigates it and gives you a report formatted for decisions.

TinyResearch: research you can make a decision on

Let's be honest, most research reports coming out of AI tools are just as much work to parse as just doing the research by hand. You get a wall of text, and you still have to dig through it for the part that answers your question before you can take any of it into a meeting. We built TinyResearch so the report is shaped around the decision: every report opens with a summary, puts comparisons and key numbers into tables you can scan, uses clear headlines, and cites every claim so you can go deeper on anything. Underneath that, there are five things we cared about more than anything else:

• Factual correctness. Every line in the report comes from a claim it pulled off a real page, with the source right there, so you can check anything you doubt.
• Credibility. Every claim is backed by a source, and how much weight it carries depends on how strong that source is for your question, which is exactly how a good researcher decides what to trust.
• Readability. A summary up top, tables for the comparisons and numbers, and clear headlines, so you see the answer up top and know where to go deeper.
• Comprehensiveness. It breaks your question into everything that needs answering, and keeps going back for another round until the gaps are closed.
• Instruction following. It sticks to the rules you give it, whether that's sources you trust, news only, research papers only or a specific date window.

A Deep report usually takes 5 to 20 minutes, returns a day's worth of analyst work, and can run in parallel. You can run up to 5 reports in parallel, and each one saves to your dashboard when it's done.

How it does on the public benchmarks

We ran TinyResearch in Deep mode on three public deep research benchmarks, using each one's own tasks and scoring. On DRACO, the benchmark Perplexity built, TinyResearch scored 51.74.

On ResearchRubrics, where experts write the requirements each report has to meet, it scores 77.10. And on DeepResearch Bench II, which grades reports against ones written by experts, it scored 46.42.

The full results, the setup and how it all works under the hood are in the TinyResearch technical blog.

TinySet: turn scattered web information into a dataset

Sometimes you already know exactly what you need: the same handful of facts across a lot of companies, projects, places or products. AI conferences happening next year with dates and ticket prices, grants for AI builders with amounts and deadlines, data centers being built in the US with operator and capacity. It's all public, it's spread across directories, company sites and individual project pages, and today someone builds it by hand in a spreadsheet, a week at a time.

TinySet builds it from one prompt. You describe what each row should be, the scope and the fields you want, review the columns it proposes, and it finds the matching entities and collects the values from public web pages. A few things make it different:

• It's made for scattered data. The harder the data is to pull together, the more TinySet is worth, because that's exactly the work nobody wants to do by hand.
• You shape it before it runs. It proposes typed columns and the one that identifies each row, and you can edit any of it, add hints on where a value usually lives, or keep only rows that pass a yes/no filter.
• It goes wide, not just deep. It splits a big ask into smaller pieces (cities, categories, listing pages) and works through each one, so you get far more than the first dozen rows.
• Any kind of data. Companies, conferences, prices, policies, places, infrastructure, anything that lives on the public web. You define the columns, so it isn't limited to a fixed set of company or people fields.
• Built from pages it actually read. Its agents only write what they found on public pages they opened, and you can add source URLs and dates as columns, so anyone reviewing it can check where a value came from and when.
• Use it as is or take it further. Watch rows land as they're collected, share a read-only link, or pull the rows into your sheet, notebook or agent through the API and MCP.

The full story of how TinySet plans and builds is in the TinySet technical blog.

A few things people are doing with them

We've been running these on everything internally, and a few have become team favorites. Each one has the real output and the exact prompt, so you can steal it.

Where is the AI data center buildout headed? A TinySet dataset of data centers under construction in the US, with operator, location and planned capacity. Alongside it, a TinyResearch report on where the buildout is heading and what's actually driving those locations.

Which humanoid robotics companies are genuinely ahead? A TinySet dataset of humanoid robotics companies with their funding, robot and stage, and a separate TinyResearch report on which ones are genuinely ahead and which are mostly hype.

Why are AI startups shutting down? A TinySet dataset of AI startups that shut down in 2026, with what they built and how much they raised, and a TinyResearch report on why they're shutting down and the patterns behind it.

Will any small modular reactors switch on this decade? A TinySet dataset of small modular reactor projects around the world with developer, country, capacity and status, and a TinyResearch report on which ones are closest to switching on and why.

What is YC betting on in AI right now? A TinySet dataset of the AI companies in Y Combinator's Fall 2026 batch, with what they build and their category, and a TinyResearch report on what YC is betting on and the themes that stand out.

There are more than 30 reports and datasets at tinyfish.ai/deep-work, all with outputs you can download and the exact prompts.

Where you can use them

Both live on the same TinyFish account and wallet as TinySearch, TinyFetch, TinyAgent, TinyBrowser and TinyMonitor, and you can hand them to the agent you already use, whether that's ChatGPT, Claude, Cursor, Codex or your own product.

• TinyResearch: the Playground, the ChatGPT plugin, the Claude connector, API, MCP, CLI and SDKs
• TinySet: the Playground, the ChatGPT plugin, the Claude connector, API, MCP, CLI and SDKs
• Pricing: A Deep TinyResearch report is $2 and a Max report is $3, and you only pay for the ones that finish. A TinySet dataset is 10 cents per row.
• Free to start: every new account gets $8 in the wallet, which is plenty to try both

What's next

This is a first release, and we know where the edges are. Both work on the public web only for now, and a TinySet dataset is a snapshot of the web at the moment you build it, so today you run it again when you want fresher data. A run collects up to 250 rows, and it can come back with fewer than you asked for, depending on what's actually on the web.

That's the first thing we're fixing on TinySet. We're working on keeping datasets updated and current on their own, so a dataset you build once stays useful instead of slowly going stale. We're also working on bigger datasets.

On TinyResearch, we're pushing on coverage, meaning finding more of what an expert would include in a report, and on making the evidence behind every conclusion even stronger. We're also adding a notification when your report is ready, so you don't have to keep the tab open.

We'd love to hear what breaks, and what you wish they did.

Go hand off the thing you've been putting off. Start at agent.tinyfish.ai, and tag @Tiny_Fish with what you asked and what came back.

We read everything.