Status · NovaRouter

Status

9 of 9 models operational

Every model NovaRouter routes to, with the state that decides whether a request to it is accepted. A disabled model stays listed rather than disappearing — a 503 you can explain beats a 404 you cannot.

Fable-5FableStreaming

Anthropic · 1M context · 128K max output

Endpoint: /v1/chat/completions

Claude-Opus-5ClaudeStreaming

Anthropic · 1M context · 128K max output

Endpoint: /v1/chat/completions

Claude-Sonnet-5ClaudeStreaming

Anthropic · 1M context · 128K max output

Endpoint: /v1/chat/completions

GPT-5.6 SolGPTStreaming

OpenAI · 400K context · 128K max output

Endpoint: /v1/chat/completions

Gemini-3 ProGeminiStreaming

Google · 2M context · 64K max output

Endpoint: /v1/chat/completions

DeepSeek-V4DeepSeekStreaming

DeepSeek · 256K context · 64K max output

Endpoint: /v1/chat/completions

Grok-4.5GrokStreaming

xAI · 256K context · 64K max output

Endpoint: /v1/chat/completions

GLM-5.2GLMStreaming

Z.ai · 200K context · 64K max output

Endpoint: /v1/chat/completions

Nova-Embed-1GLMEmbeddings

Z.ai · 8K context · 1024 dimensions

Endpoint: /v1/embeddings

No availability history

NovaRouter does not probe upstreams on a schedule, so there is no uptime percentage to report — health is discovered when a request fails and moves to the next candidate in that model's chain. Every attempt is recorded against the request it belongs to, so failover is visible per request on your own usage page. Earlier versions of this page showed a 90-day uptime figure that was never measured; it has been removed rather than estimated.

Publisher network

6 publishers, one wire format

Every model in the catalogue answers the same OpenAI-compatible request. Each alias carries an ordered chain of upstreams, and a rate limit or 5xx from one moves to the next without a retry in your code.

Anthropic

us-east • us-west • eu-central

Operational

Fable and Claude, with the longest context windows in the catalogue and the strongest agentic coding scores.

3modelsEnterprise

OpenAI

us-east • eu-west • apac

Operational

GPT with broad general capability, native structured output and audio input.

1modelEnterprise

Google

us-central • eu-west • apac

Operational

Gemini, with a two-million-token context window and native multimodal input across text, image and audio.

1modelEnterprise

DeepSeek

apac • us-west

Operational

Open-weight reasoning models priced near the floor, strongest on mathematics and competitive programming.

1modelScale

xAI

us-central • eu-west

Operational

Grok with fast reasoning over a wide context window and real-time knowledge access.

1modelScale

Z.ai

apac • us-west

Operational

Open-weight GLM served at the lowest cost per token, plus the catalogue's embedding model.

2modelsScale

Models per publisher

Anthropic3
Z.ai2
OpenAI1
Google1
DeepSeek1
xAI1

How routing works

Ordered chains per model
Every alias declares its upstream candidates in order. Position 0 is tried first.
Failover without a retry in your code
A rate limit, timeout or 5xx moves to the next candidate. A malformed request stops immediately — replaying it would fail identically.
Bounded, then honest
One chain walk is bounded end to end. If no candidate answers, the response is a 503 that says so rather than a hung connection.

What a failure looks like

Each attempt is recorded against the request, so a completion that took three tries is visible as three attempts rather than as one slow call.

Only the attempt that succeeded contributes tokens to your usage record — a request nobody could serve records none at all.