Septor Labs

Septor is a managed gateway for agent workloads. Each call goes to a provider that is still answering. You set the policy: priority, a weight split, least connections, or latency. A circuit breaker pulls a failing host out of rotation.

If that provider dies before the first token, the call fails over. The agent keeps going, and you are not billed for a response that never started.

Repeat prompts can hit an opt-in semantic cache. A cache hit is billed at 70% of the completion.

Credits are prepaid. Each key has its own budget. There is no per-seat fee.

Pricing is not a single upstream. Range-priced models publish a floor and a ceiling, and a lot of them show a cheapest rate 30% or more under that ceiling. The call is billed at the provider that served it, and it never goes above the ceiling.

OpenAI chat clients and Claude-style clients on the Anthropic messages path use one router. Set the model to auto, or pin a model id.