Introduction

Eterial is an inference API that speaks the OpenAI protocol. Requests are served by gonka.ai, a decentralised GPU network, and any request the network cannot complete falls back automatically to the same model on OpenRouter. You keep the SDK you already use, the request bodies you already send, and the response shapes you already parse.

Reliability

There are two numbers here, and only one of them is yours.

The network's. Gonka is early technology, and it drops requests: depending on conditions, between 10% and 35% of them fail to complete on the network. Taken alone, that would rule it out of production.

Ours. Every one of those failures is retried against the equivalent model on OpenRouter before anything is returned to you, so a request the network drops is not a request that failed. Send inference through Eterial and it gets served — the network going down is something we absorb, not something you see.

That is what makes a decentralised network usable behind a real workload today. You do not have to write the fallback yourself, watch the network's health, or keep a second provider wired up for the days it is struggling — that work is ours, and it happens inside the request you already sent.

What makes it different

• One rate for input and output (for requests the network serves). Most gateways charge several times more for generated tokens than for prompt tokens. On the network the two rates are identical, so the cost of a request is simply its total token count times a single number. A request that falls back to OpenRouter is priced the way OpenRouter prices it — input and output separately.
• Open-weight models at open-weight prices. Served without the margin a closed model carries. The catalogue lists what is available today; GLM-5.2 Coming soon is next.
• Modalities the network does not have. Because the fallback provider supports more than Gonka does, the same API offers capabilities the network cannot serve at all. Images and PDFs go in as ordinary OpenAI content parts, and the request is routed to a backend that can read them — see Multimodal content. Web search works the same way, and is run here rather than by the model: Web search.
• A one-line migration. Point base_url at Eterial and pass an Eterial key. Everything else in your code stays as it is.

Where to start

Reading these docs as a model

Every page here is available as plain markdown as well as HTML:

• append .md to any docs URL — /docs/quickstart.md
• /llms.txt lists every page with a one-line summary
• /llms-full.txt is the entire documentation set in a single file