StarSeaAPI
Platform
A relay is only worth using if it is cheaper, more available, or easier than going direct. Here is what this platform does on each of those three axes — and what it does not claim.
Multi-provider supply, no single point of failure
Every model is served through more than one independent upstream route. A provider incident, a quota exhaustion or a regional failure degrades capacity rather than availability, and traffic is re-routed without intervention.
Cache-aware routing that lowers the real cost of long sessions
Upstream model providers bill repeated prompt prefixes at a discount, but only if the request lands on the node that already holds the cache. The router anchors a conversation to the same upstream for as long as it runs, so agent loops and long documents keep hitting cache instead of re-paying prefill on every turn.
Capability bridging closes the gaps between providers
Suppliers are uneven: some models are strong but text-only, some expose only one endpoint style, some speak a protocol your client does not. The gateway closes those gaps — images can be sent to text-only models, image/video/speech generation is presented through one contract, and three protocol styles are terminated natively.
Reliability designed in, not patched in
Health checks run on short and long windows so that spikes are not mistaken for failures and slow degradations are not missed. Deployments ship as a parallel environment validated before traffic moves, with the previous version kept as a fallback path, so releases do not interrupt in-flight requests.
Pricing you can audit
Every model page shows the input rate, the output multiplier and, for image models, the per-request price. Cache hits are metered and billed at the discounted rate rather than silently averaged, and per-key usage can be checked without contacting anyone.
Self-serve from first call to production
The model catalogue tracks upstream availability automatically, a one-click import script configures your local clients, and there is no seat licence, sales call or minimum commitment between signing up and shipping.
Models
61 models are currently on sale, each with an individual page listing price, measured benchmarks and setup.
What we do not claim
No platform is free of trade-offs. Model behaviour still changes when an upstream updates a model, so dated variants are exposed as separate ids. Some long-tail models are served by a single upstream and are therefore less resilient than the main families. And answers are only as good as the model you choose — routing cannot fix a model that is wrong for the task.