Routing & failover
Routing & failover
You only choose the model; ZeroToken takes care of getting the request to a healthy upstream.
Multi-account scheduling
A model is usually backed by several upstream accounts. For each request the gateway picks an account based on its health, current load and rate-limit state.
Automatic failover
If an upstream account returns a rate limit, overload or transient error, the gateway retries on another account transparently:
Only the request that finally succeeds is billed; failed attempts in between are not; An error is returned to you (usually 429 or 5xx) only when every available account has failed.
Session stickiness
For long conversations and coding agents that rely on prompt caching, consecutive requests from the same session are routed to the same upstream account whenever possible, improving cache hits and lowering cost.
Composite groups
Composite groups choose the upstream platform from the model name. Requesting deepseek-chat goes to DeepSeek, claude-sonnet-4-5 goes to Anthropic — one key, many vendors.
Recommendations
Retry 429 and 5xx responses with exponential backoff; Use streaming for long tasks so proxies and CDNs don't time out idle connections.