Always one frontier model
Simple. Quality is consistent. The bill is also consistent — everyday refactors and 'ok' turns pay frontier rates. No cross-client memory unless the editor ships its own.
You can send every turn to one frontier model. You can build your own retrieval layer. Both are valid. Strix exists when you want lanes, filtering, and account memory without standing up that stack.
Simple. Quality is consistent. The bill is also consistent — everyday refactors and 'ok' turns pay frontier rates. No cross-client memory unless the editor ships its own.
Strix memory ships in the gateway you already call: silent extract and inject, project scopes, pin/CRUD, recall, history, feedback, batch, MCP, and forget on the same strix_ key. Dedicated retrieval products remain useful for graphs, connectors, and document corpora.
You pick models per request. Fine if you already classify work yourself. Strix is the classifier, filter, and memory layer in front of that gateway. See vs OpenRouter.
You want a single vendor model with no router. You need a full-document RAG corpus, not short derived facts. You will not store an OpenRouter key. You need us to host inference without BYOK — that is not the default product.
Still comparing? Compare index · OpenRouter · Cursor · Memory SDKs · Memory · Pricing
Trial needs an OpenRouter key. Tokens stay on their bill. Strix is the router.