Guide · from the data

What AI products are built with: models, frameworks and data, by the numbers

The model APIs, frameworks, databases and hosting behind 987 makers' AI products we track — and where the momentum is moving.

Published
September 30, 2026
Numbers refreshed
November 3, 2026
Reading time
3 min
Share
Post on X

We track 987 makers' products that show which model API they call — from the maker's own words, a subprocessor list, a customer story or the product's website. That's the sample below: live AI products, from indie launches to funded startups.

OpenAI is the default; almost a quarter use more than one model

OpenAI's API appears in 736 of them, Claude in 252, Gemini in 99. But 203 products (21%) use two or more model providers directly, and 89 reach models through a gateway such as OpenRouter or LiteLLM. Picking one model forever is no longer the norm. The usual reasons: cheap requests go to cheap models, and a second provider covers outages or tasks the first does worse.

The momentum isn't where the installed base is. Against the typical package we track, Claude's SDK downloads grew 4.4× as fast over the last year, Gemini's 4.9×, OpenRouter's 5.1× — and OpenAI's 1.6×.

Frameworks: LangChain has the installed base, the TypeScript tools have the growth

Among the 42 AI products that show an agent framework, LangChain dominates. The download trend says something else: LangChain's grew 1.0× as fast as the median package — in line with the registry — while Mastra's grew 3.2× and the Vercel AI SDK's 2.6×. The AI SDK is downloaded millions of times a week but rarely visible from outside a product: it's a library, so it shows up in code, not on websites.

The rest of the stack looks like any web product

For their database, the 283 AI products that show one mostly chose a backend platform or managed Postgres:

And they host where web apps host — Vercel first, then Amazon Web Services:

Vector search and observability: visible only when they're big

Only 40 of the 987 show a vector database, and 63 an LLM observability tool. That's far below how many use them: both run behind the product, where no website check can see them, and makers rarely mention them. Among those we can see, Pinecone leads vector search — but its SDK's downloads grew 0.7× as fast as the median package, while pgvector's grew 1.3×, which suggests more builders keep vectors next to their data in Postgres.

A tool missing from these counts isn't unused — it's unobserved. Each tool's page shows its usages by kind of source.

What this means if you're starting now

  • Put model calls behind one function or an SDK that switches providers; a fifth of these AI products already run more than one model.
  • A gateway (OpenRouter, LiteLLM) is worth it once you use two providers or need fallbacks.
  • Start vector search in Postgres (pgvector) unless you have millions of vectors or need heavy hybrid search.
  • Add LLM tracing before launch. It can't be seen from outside, so the counts here say little about who uses it — but it's the only way to see what each user costs you.