What Am I Actually Choosing Between? We Sorted 1,347 AI Tools by Shape

Somebody asks what the difference is between ChatGPT, Copilot, Claude Code and one of the agent platforms that turned up last quarter, and the answer usually arrives as a tidy list of categories. Chat assistants over here. Copilots that live in your software. Command-line agents. Autonomous runtimes. Model providers. Five boxes, one product dropped into each, question closed.

We were going to write that list. Then it occurred to us that we are sitting on something that could test it. Our AI directory holds 1,347 active entries, each a researched description of a real product. A list of categories is a claim about a population, and a claim like that can be tested by counting.

So we did, and the tidy list did not survive it. The biggest thing in the market has no box in that list at all, and the box everyone pictures first — the chat assistant — turns out to be about one entry in forty.

There are two questions here, and only one of them our directory answers

Our directory sorts everything into twelve categories: Code & Dev Tools, LLMs & Chat, Video & Audio, Business & Finance, Productivity, Agents & Automation, Image Generation, Research & Data, Security & Safety, Education, Healthcare, Other. Every one of them answers the same question — what does this thing make?

That is a useful question, but it is not what somebody means by "what am I actually choosing between". They are asking about shape. Is this a thing I talk to? A thing that lives inside software I already pay for? A thing I hand a goal and walk away from? Or a part I would use to build something else? Nothing in our schema records any of it. Guide #16, on choosing a tool without believing the marketing, had already noted that the directory has no field for which model runs underneath a product or for what happens to your data. This is a third missing field: what kind of thing it is.

So we added a field by hand and sorted products into six shapes. The definitions turn on who decides the work, not on what comes out.

• Single-purpose application. The product decides the job; you bring an input and it returns one kind of output. LALAL.AI splits a track into stems, Photomath reads a photographed equation, and you do not ask either of them for anything else.
• General chat assistant. You bring the task, turn by turn, and the range is open. ChatGPT.
• Attachment. Its primary home is inside somebody else's product. Notion AI is not a destination. It is a thing Notion has.
• Agent runtime. You state a goal and it takes several steps without you in each one — Aider in a terminal, Factory turning a ticket into a pull request.
• Building block. Parts sold to developers: weights, model APIs, inference, frameworks, the plumbing agents run on. See Ollama, AWS Bedrock, Browserbase.
• Not a product. It is in the directory and you cannot choose it.

Then we took a sample: 200 of the 1,347, picked by hashing each entry's slug and taking the lowest 200 hashes. That is deterministic — anyone re-running it gets the same 200 — and it is uncorrelated with the order entries were added, which matters because that order tracks category. Each of the 200 was read and labelled from its own name, category, tagline and description. Nothing else.

What we tried first, and why it did not work

The obvious approach was to skip the reading and sort all 1,347 with a keyword rule. We wrote one. It used a lexicon per shape, matched on word boundaries across name, tagline, description, tags, use cases and deliverables, and deliberately excluded the adjectives that describe nothing.

Scored against the 200 hand labels, it agreed 23.5% of the time. It returned no label at all for 39% of them and two or more labels for 16%. Where it disagreed, the reason is the same reason the question is hard: the words have come loose from the things.

Browserbase says it is "managed-browser infrastructure for AI agents" and the rule filed it as an agent — it is infrastructure that agents are pointed at, which is the opposite end of the transaction. Glean says "AI assistant" and is enterprise search. Grammarly says "across every app you write in" and the rule read that as a plugin, which is closer than most of its guesses. In 2026, "AI agent" and "AI assistant" mean about as much as "cloud" did in 2013. They are phrases a marketing department attaches to a product regardless of the product's shape, and a classifier built on them measures the marketing.

We kept the rule and its score in the guide rather than dropping it, because "we could not do this automatically" is itself the finding a reader can use. If a keyword scan of the description cannot tell you what a thing is, neither can a skim of the homepage.

The distribution

General chat assistants are 2.5% of the corpus. Five entries in two hundred. The interval runs from 0.5% to 4.5%, so the true figure across all 1,347 is somewhere between about seven products and about sixty — and every reading in that range says the same thing. The category that defines the whole subject in most people's heads is a rounding error in the population.

The shape nobody names is the one that dominates: at 52.0%, the single-purpose application — a product that does one job and only that job. Behind it, at 27.5%, sit the building blocks, parts rather than products, sold to somebody building something else. Attachments are 8.5%, agent runtimes 7.5%.

Line that up against the five-box list and the arithmetic is unkind. Chat assistant, copilot, command-line agent, autonomous runtime and model provider between them describe 46% of what is out there. The other 54% is single-purpose applications plus a small residue of things that are not products. The list is not wrong, just incomplete. It is missing the box most of the market sits in.

The category named after chat is mostly not chat

Cross the two axes and you can see where a browse goes wrong. Of the 34 sampled entries filed under LLMs & Chat — the category anyone hunting for a chatbot would open first — 26 are developer building blocks. Open-weight model families, inference clouds, gateways. Four are chat assistants. Somebody arriving there wanting something to talk to is mostly being shown engine parts.

The reverse holds too, and more cleanly. Image Generation is 20 single-purpose applications out of 21; Video & Audio, 14 of 15; Healthcare and Other, all of them. Those categories are nearly pure on both axes, because there a product's function and its shape are the same thing.

Code & Dev Tools is the one category that splits four ways at once — 14 building blocks, 8 agent runtimes, 6 attachments, 4 single-purpose apps out of 32. A developer opening it is handed four different kinds of thing under one heading, with no filter to separate them.

So what do you actually ask?

Four questions, in this order. They are what the labelling rubric turned into, and each is answerable from a product's own page in about a minute.

Who decides the task? If the product decides — it transcribes, it upscales, it drafts a demand letter — you are buying a single-purpose application and the only questions that matter are whether it does that job well and what it costs. Half the market is here, and it is the half where evaluation is easiest.

Where does it live? An attachment inside software you already pay for is a different purchase from a destination you have to remember to visit. It also tends to be a different budget line and a different data-sharing question.

How much happens while you are not watching? This is the one that escalates. A chat assistant can do nothing on its own; you read every reply before anything happens. An agent runtime reads, decides and acts in a loop, and the reading includes whatever text it encounters along the way. That is where the prompt-injection problem stops being a curiosity and starts being money — the same argument as in the vocabulary guide, one notch louder. Before you let anything run unattended, know what it can touch — and where what it touches ends up.

Who is the customer? If the answer is "a developer", you are looking at a part. Cerebras Inference is not an alternative to ChatGPT in any sense that helps you; it is somewhere a chat product might get its tokens from. A quarter of the directory is parts, and a lot of comparison articles quietly mix them in with products.

What auditing the directory turned up

Reading 200 entries closely does things to your opinion of your own database. Two defects came out of it, and both are recorded here. A guide about counting your options should not go quiet about its own miscounts.

Nineteen products were listed twice. We clustered every active entry by website host and by product name normalised for case, punctuation, bracketed suffixes and vendor filler. Thirty-eight candidate pairs came back; nineteen were the same product entered twice, both indexed and both returned by our own search. Superhuman and Superhuman AI. Granola and Granola Notes. Codeium and Windsurf, which are one product under two of its three names. Two of the nineteen said so in their own taglines and were indexed anyway: Hume EVI read "(subset of Hume AI)", and Amazon CodeWhisperer read "Rebranded as Amazon Q Developer in 2024 — same tool, new name."

The other nineteen candidates are real siblings — Cursor and Cursor Composer, Suno and Suno Studio, Topaz Photo AI and Topaz Video AI — and were left alone. The duplicates are now out of Google and out of our search. Nothing was deleted and no URL stopped working.

Update, later the same day. This guide first went out saying we had not merged the writing — that in every pair both entries carried separate research not contained in the other, so picking which sentences survive was an editorial job we were recording as owed rather than doing. It has since been done: all nineteen merged by hand, each survivor rewritten to carry what its retired twin knew, and each one cut back inside the 250-to-350-word limit our own entry standard sets, because straight concatenation would have produced 500 to 700 words and failed it.

Doing it turned up something the retirement alone had missed, and it sharpens the point of this whole section. Reading both halves of each pair side by side found seven places where the two records of the same product contradict each other. Glean's two entries date the same funding round eighteen months apart. Filmora's two carry different product line-ups on the same review date. Captions' two disagree about whether its customer support is good. And Recorded Future's two disagree about who its customers are: one says the platform serves "45 of the Fortune 100", the other says 45 national governments and more than half the Fortune 100. The 45 attached itself to the wrong noun somewhere, and it sat there being read until two copies of the same entry were put next to each other.

A duplicate entry is not wasted space. It is two divergent accounts of the same thing, and you believe whichever one search happens to serve you.

Fixing that broke our own drift checker, which is how we found the second defect. The tool that compares the database against the search index promises in its own comments that each of its counts mirrors the rule the model uses to decide what gets indexed. For directory entries it did not: the model tests three conditions, the checker tested one. That gap was invisible for exactly as long as the other two conditions were universally true — and the moment nineteen rows used one of them, the checker reported nineteen rows of drift against an index that was perfectly correct. A rule that only agrees with reality while a column sits unused was never really mirroring anything. Corrected, with a test that fails against the old version.

Most of what we found was not broken. Dead products are handled well: of 36 entries whose own tagline says the thing is shut down, sunsetting or discontinued, most are already hidden from search, and the ones still visible are open-weight models that remain downloadable. The entry standard held — 1,267 indexed entries pass it with zero errors. We went looking for rot and mostly did not find it.

Four of the 200, though, are things you cannot choose. Safe Superintelligence is a research lab that says openly it has no commercial product. Lemonade Insurance AI is an insurance company. BloombergGPT is a model Bloomberg confirmed is not used in any live tool. Google Sec-PaLM is branding Google has retired. All four are interesting and none is a thing a reader can go and use, which is a category question our schema cannot express either.

Where this stops

Every number here comes from our own database. We did not install anything, test any product, or fetch a single vendor page for this piece — so if one of our entries has drifted out of date, our label for it inherits the error. The risk is real and roughly measurable. 87.7% of the active corpus carries a review date, so about 166 entries have a description that has never been re-checked since it was written.

Two hundred is a sample, and the confidence intervals on the figure are the honest width of what it supports. They are not narrow. "Attachments are 8.5%" really means somewhere between 4.9% and 12.1%. What the sample supports firmly is the ordering and the size of the gaps, not the second decimal place.

And the labels are judgements, not measurements. The hard ones were hard in a consistent way — they were products whose shape is a chat window but whose scope is a single job. Replika is a conversation, but only one kind of conversation. AI Dungeon is a chat interface wrapped around a text adventure. Both were labelled single-purpose, on the grounds that the product decides the job, and a reasonable person could file them the other way. If every borderline case flipped, chat assistants would still be under a tenth of the corpus.

One more thing you should weigh. Claude Code, Claude Computer Use and the Anthropic API are entries in the corpus we counted, and this guide was drafted with an Anthropic model. We have kept named products to worked examples rather than recommendations for that reason, and the labelling rules are published above so you can disagree with a specific call rather than with our motives.

If you want to browse the population yourself rather than take the summary, the directory index is the place to start, and the guide on evaluating a tool once you have found it is the natural next step. It will tell you what to check. This one only tells you what kind of thing you are looking at — which, it turns out, most people get wrong before they start.