OpenAI settings
The key and the models the project uses.
Why the project is built with Claude models while the key here is OpenAI
These are two different things, and they do not contradict each other. Your project's code is written by a developer on their own machine, in Claude Code, with Claude models — Sonnet, Opus, Fable. The key on this page is not for development: it powers the services that run ON YOUR SERVER and answer your visitors — the Quiz, vector search, the knowledge graph.
The decisive reason is technical, not a matter of taste. Vector search and the knowledge graph rest on text embeddings. Anthropic does not make them: its documentation states verbatim, «Anthropic does not offer its own embedding model», and points to a third-party provider. A product that needs search by meaning must take embeddings where they exist.
The second reason is about money, and it is about keeping two budgets apart. A Claude subscription is issued to a PERSON: it is an allowance for their own work, shared with chat and running on rolling windows. A server answering visitors is a different kind of load — continuous, and growing with traffic. That belongs on a metered API key, not on a developer's working allowance: otherwise the first surge of visitors eats the very thing the project is built with.
The key is yours, the cost goes straight to your provider, and we never see it. The list of providers is not carved in stone: the moment the services we use gain another suitable option, it will appear here too.
OpenAI API key
What draws on this key
The key is stored only on your own server, in the services' own environment files. It is never sent anywhere else and never shown back to this page — what you see is a state, not the value.
One key — what spends it, and why this is not a subscription
What you get. One key that every part of the platform draws on. You are charged for what was actually used, down to the request, instead of paying a flat fee for capacity you may not touch.
What spends it. Vector memory turns each chunk of text into a vector — once at ingest, once per search. The knowledge graph spends far more: it reads every chunk to extract entities and relations, then writes an answer for each question. Voice input pays for transcription by the minute; images pay a vision model by the picture. Different jobs, different models, different prices — a search costs a fraction of a cent, a graph pass over a long document costs minutes of compute.
Why a subscription cannot replace it. No single flat plan covers this shape of work: embeddings, transcription, vision and reasoning are separate services priced separately. More to the point, most consumer plans restrict commercial use and many forbid automated or agent-driven access outright — which is exactly what a platform like this does all day. A pay-as-you-go key is not a workaround; it is the only arrangement that is both permitted and predictable.
What we do about the cost. Tokens are treated as money, because they are. Embeddings run on the cheap text-embedding-3-small rather than the large model — the quality difference is imperceptible for this work and the price is about seven times lower. The vector store answers from a real index, so a search reads a handful of rows instead of the whole base. The graph pays its expensive pass once at ingest and stays cheap per question afterwards.
Where to watch it. Loading a large corpus into the knowledge base is the one operation that costs real money in a single go. Load what you will actually ask about, and remember that re-loading a document pays for that pass again.
Changing the provider. OpenAI is the default chosen for the balance of price, quality and availability. If your project needs a different provider or model, write to admin@fractera.ai