One line stands between your code and every model.

Your OpenAI client keeps working - same SDK, same request shape, same response shape. Only the base URL changes. Behind it, every request lands on the cheapest healthy upstream serving that model and fails over on its own. Tokens bill at the rate the vendor invoices us; the platform fee is charged once, when you buy credits.

The migration is one line

Your existing OpenAI client keeps working - same SDK, same request shape, same response shape. Point it at https://api.blazerail.com/v1.

What happens to one request

An identical request inside the 5-minute window is served from Redis, and is not billed for tokens. Otherwise upstreams serving that model id are ordered by recorded price and measured latency. A failed call is retried on the next healthy upstream serving the same model id - never a substitute model - and the failure is recorded against the upstream that caused it.

What the memory graph holds

Facts extracted from conversations are linked to the conversations they came from, to the facts they supersede, and to the facts they relate to, and the conversations themselves are chained in the order they happened. Four link types are recorded: sourced from, supersedes, relates to, and followed by. The diagram on the page is illustrative - in it, a fact learned in a Monday session on gpt-5.6-sol is superseded by one learned in a Saturday session on deepseek-v4.

Memory and document extraction

Memory is included, retrievable by vector, full-text and knowledge graph, and scoped across sessions, keys and models. Document extraction runs on PaddleOCR and returns text plus per-page confidence.

The whole fee schedule

Credits carry one platform fee, charged when you buy them. Usage then bills at the rate the model’s page publishes, with no percentage added on top. A response served from cache carries no token charge at all, just a $0.00005 infrastructure fee for the request.